mamba心脏疾病诊断

1. medmamba复现

想首先尝试一下medmamba的实验复现，用做后续的基础架构

是二维的，不好改，还是改改之前的分割网络看看

2. echo-mi实验

2.1 idea

使用双切面（二腔室+四腔室）超声心动图数据训练Mamba模型诊断早期心梗的需求，设计了一个创新性的双路径时空融合Mamba架构（Dual-Path Spatiotemporal Fusion Mamba, DPSF-Mamba）。该架构针对性解决多切面心脏超声的时空特征融合问题，核心思路如下：

一、核心改造思路

双路径异构特征提取

独立编码路径：为二腔室（2C）和四腔室（4C）切面分别设计专用Mamba块，适应不同视角的局部结构特征。
2C路径：聚焦左心室前壁、心尖部运动异常（早期心梗敏感区域）。

4C路径：捕获室间隔、侧壁运动及整体心室协调性（诊断关键指标)。

动态门控融合模块（DGFM）：引入可学习的门控权重，自适应融合双路径特征（公式示例）：

$$F_{fused} = \sigma(W_g \cdot [F_{2C}, F_{4C}]) \odot F_{2C} + (1-\sigma(W_g \cdot [F_{2C}, F_{4C}])) \odot F_{4C}$$

其中 $W_g$ 为可训练权重，$\sigma$ 为Sigmoid函数，实现特征重要性动态分配。

多任务联合优化分类头

//todo

二、创新技术优势

模块	传统Mamba局限	DPSF-Mamba改进	临床价值
多切面处理	单一路径忽视视角差异	双路径异构编码 + 动态融合	减少视角偏差，提升小病灶敏感性
时空特征利用	长序列建模但空间关联弱	时空切片重组 + 坐标注意力	同步捕捉运动异常与结构变形
数据效率	需大量标注数据	多任务学习共享特征	缓解超声标注稀缺问题
可解释性	黑盒决策	门控权重可视化切面贡献度	辅助医生理解AI诊断依据

三、预期效果验证方案

可解释性分析

绘制门控权重热力图（如下示例），验证模型对病变切面的关注度：

前壁梗死患者：4C切面权重峰值达0.83（主导诊断）
心尖梗死患者：2C切面权重升至0.79

四、潜在挑战与解决方案

挑战1：双切面数据不全（如部分患者缺失一个切面）
方案：引入跨切面知识蒸馏，用完整数据训练教师网络指导单切面学生网络。
挑战2：超声伪影干扰
方案：在Mamba前端加入对抗去噪模块（如Conditional GAN）。

2.2 faec_advance实验

name	acc	auc	f1	precision	recall	specificity
faec	0.71875	0.7013	0.3077	1.0	0.1818	1.0

camus 多切面

name	acc	auc	f1	precision	recall	specificity
Experiment_6 fold_1	0.8125	0.858333	0.769230	0.714286	0.833333	0.8
Experiment_6 fold_2	0.75	0.741666	0.555555	0.833333	0.416666	0.949999
Experiment_6 fold_3	0.8125	0.899999	0.699999	0.875	0.583333	0.949999
Experiment_6 fold_4	0.78125	0.852814	0.588235	0.833333	0.454545	0.952380
Experiment_6 fold_5	0.71875	0.701298	0.307692	1.0	0.1818	1.0
Experiment_6 avg	0.775	0.8108	0.5841	0.8512	0.4939	0.9305

3. MIMamba实验

改造之前的分割网络的mamba用于分类

3.1 camus数据集

camus A2C的实验

name	acc	auc	f1	precision	recall
Experiment_4 fold_1	0.375	0.533333	0.545454	0.375	1.0
Experiment_4 fold_2	0.34375	0.558441	0.511628	0.34375	1.0
Experiment_5 fold_1	0.75	0.699999	0.636364	0.699999	0.583333
Experiment_5 fold_2	0.71875	0.670995	0.571428	0.6	0.545454
Experiment_5 fold_3	0.71875	0.636363	0.470588	0.571428	0.4
Experiment_5 fold_4	0.71875	0.590909	0.470588	0.571428	0.4
Experiment_5 fold_5	0.625	0.490909	0.4	0.4	0.4
Experiment_5 avg	0.70625	0.617835	0.509794	0.568571	0.425757

camus 多切面

name	acc	auc	f1	precision	recall	specificity
Experiment_7 fold_1	0.84375	0.837499	0.761904	0.888888	0.666666	0.949999
Experiment_7 fold_2	0.75	0.708333	0.555555	0.833333	0.416666	0.949999
Experiment_7 fold_3	0.8125	0.8375	0.727272	0.8	0.666666	0.899999
Experiment_7 fold_4	0.8125	0.818181	0.75	0.692307	0.818181	0.809523
Experiment_7 fold_5	0.71875	0.714285	0.307692	1.0	0.181818	1.0
Experiment_7 (监控acc)	0.7875	0.7832	0.6205	0.8429	0.5500	0.9219
Experiment_8 fold_1	0.65625	0.875	0.266666	0.666666	0.166666	0.949999
Experiment_8 fold_2	0.75	0.816666	0.666666	0.666666	0.666666	0.8
Experiment_8 fold_3	0.71875	0.858333	0.571428	0.666666	0.5	0.85
Experiment_8 fold_4	0.65625	0.8658	0.0	0	0	1
Experiment_8 fold_5	0.6875	0.796536	0.583333	0.538461	0.636363	0.714285
Experiment_8 (监控auc)
Experiment_9 fold_1	0.75	0.879166	0.5	1.0	0.333333	1.0
Experiment_9 fold_2	0.65625	0.695833	0.153846	1.0	0.083333	1.0
Experiment_9 fold_3	0.65625	0.770833	0.153846	1.0	0.083333	1.0
Experiment_9 fold_4	0.6875	0.649350	0.166666	1.0	0.090909	1.0
Experiment_9 fold_5	0.75	0.761904	0.5	0.8	0.363636	0.952380
Experiment_9 (precision)
Experiment_10 fold_1	0.84375	0.875	0.814814	0.733333	0.916666	0.8
Experiment_10 fold_2	0.78125	0.75	0.695652	0.727272	0.666666	0.85
Experiment_10 fold_3	0.84375	0.85	0.761904	0.888888	0.666666	0.949999
Experiment_10 fold_4	0.75	0.779220	0.692307	0.6	0.818181	0.714285
Experiment_10 fold_5	0.625	0.675324	0.625	0.476190	0.909090	0.476190
Experiment_10 (监控f1)	0.7688	0.7859	0.7179	0.6851	0.7955	0.7581
Experiment_11 fold_1	0.90625	0.85	0.869565	0.909090	0.833333	0.949999
Experiment_11 fold_2	0.75	0.754166	0.714285	0.625	0.833333	0.699999
Experiment_11 fold_3	0.78125	0.820833	0.740740	0.666666	0.833333	0.75
Experiment_11 fold_4	0.75	0.744588	0.714285	0.588235	0.909090	0.666666
Experiment_11 fold_5	0.65625	0.679653	0.645161	0.5	0.909090	0.523809
Experiment_best	0.8063	0.8129	0.7321	0.7512	0.7242	0.8547
Experiment_15 fold_1	0.875	0.870833	0.846153	0.785714	0.916666	0.85
Experiment_15 fold_2	0.8125	0.762499	0.769230	0.714285	0.833333	0.8
Experiment_15 fold_3	0.84375	0.854166	0.761904	0.888888	0.666666	0.949999
Experiment_15 fold_4	0.75	0.709956	0.6	0.666666	0.545454	0.857142
Experiment_15 fold_5	0.75	0.718614	0.692307	0.6	0.818181	0.714285
Experiment_15
Experiment_18 fold_1	0.78125	0.725	0.72	0.6923	0.75	0.8
Experiment_23 fold_1	0.78125	0.7625	0.6666	0.7777	0.5833	0.8999
Experiment_24 fold_1	0.8125	0.8166	0.6999	0.875	0.5833	0.949999
Experiment_25 fold_1	0.84375	0.862499	0.761904	0.888888	0.666666	0.949999

使用adaw

name	acc	auc	f1	precision	recall	specificity
Experiment_28 fold_1	0.8125	0.791666	0.75	0.75	0.75	0.85
Experiment_29 fold_1	0.8125	0.7958	0.75	0.75	0.75	0.85

camus mi_mamba A2C和A4C路径

d_state: 16   d_state: 32

name	acc	auc	f1	precision	recall	specificity
Experiment_26 fold_1	0.78125	0.766666	0.72	0.692307	0.75	0.8

换用 adaw

name	acc	auc	f1	precision	recall	specificity
Experiment_27 fold_1	0.8125	0.75	0.75	0.75	0.75	0.85

camus 多切面 MultiStageFusion

name	acc	auc	f1	precision	recall	specificity
Experiment_12 fold_1	0.8125	0.866666	0.699999	0.875	0.583333	0.949999
Experiment_12 fold_2	0.75	0.720833	0.666666	0.666666	0.666666	0.8
Experiment_12 fold_3	0.78125	0.75	0.631579	0.857142	0.5	0.949
Experiment_13 fold_1	0.8125	0.754166	0.699999	0.875	0.583333	0.949999
Experiment_14 fold_1	0.8125	0.891666	0.785714	0.6875	0.916666	0.75

camus 多尺度融合模型

name	acc	auc	f1	precision	recall	specificity
Experiment_17 fold_1	0.71875	0.695833	0.666666	0.6	0.75	0.699999

camus 分层融合模型mi_mamba_hierarchical_model

name	acc	auc	f1	precision	recall	specificity
Experiment_16 fold_1	0.78125	0.8208	0.7407	0.666666	0.833333	0.75
Experiment_16 fold_2	0.6875	0.5208	0.375	0.75	0.25	0.9499
Experiment_16 fold_3	0.75	0.8	0.6	0.75	0.5	0.8999
Experiment_16 fold_4	0.75	0.6883	0.6	0.666666	0.545454	0.857142
Experiment_16 fold_5	0.6875	0.580086	0.375	0.6	0.272727	0.904762
Experiment_19 fold_1	0.75	0.758333	0.5555
Experiment_20 fold_1	0.75	0.7333	0.6666	0.666666	0.666666	0.8
Experiment_21 fold_1	0.78125	0.75	0.695652	0.727272	0.666666	0.85
Experiment_22 fold_1	0.75	0.6875	0.5555	0.8333	0.416666	0.9499
	0.75	0.6958	0.6	0.75	0.5	0.899999
换用adaw
Experiment_30 fold_1	0.78125	0.708333	0.666666	0.777777	0.583333	0.899999

cross-ssm

name	acc	auc	f1	precision	recall	specificity
Experiment_32 fold_1	0.625	0.475	0	0	0	1
Experiment_33 fold_1	0.78125	0.783333	0.666666	0.777777	0.583333	0.899999
Experiment_34 fold_1 gate	0.75	0.754166	0.636363	0.699999	0.583333	0.85
Experiment_35 fold_1 gate w	0.8125	0.804166	0.727272	0.8	0.666666	0.899999
Experiment_36 fold_1 attn w	0.75	0.720833	0.6	0.75	0.5	0.899999
Experiment_37 fold_1 gate attn w	0.8125	0.825	0.727272	0.8	0.666666	0.899999
Experiment_ fold_1 attn gate w	0.75	0.774999	0.636363	0.699999	0.583333	0.85

从上面的来看cross-ssm效果还可以，继续使用cross-ssm 验证camus填充方式

cross-ssm gate attn w

name	acc	auc	f1	precision	recall	specificity
Experiment_39 fold_1 repeat	0.78125	0.729166	0.666666	0.777777	0.583333	0.899999
Experiment_40 fold_1 interpolate	0.84375	0.845833	0.8	0.769230	0.833333	0.85
Experiment_40 fold_1 random_repeat	0.75	0.754166	0.666666	0.666666	0.666666	0.8
Experiment_41 fold_1 cyclic	0.8125	0.791666	0.75	0.75	0.75	0.85
Experiment_42 fold_1 reflect	0.71875	0.745833	0.666666	0.6	0.75	0.699999
Experiment_43 fold_1 noise	0.75	0.745833	0.692307	0.642857	0.75	0.75
Experiment_44 fold_1 random	0.8125	0.783333	0.727272	0.8	0.666666	0.899999

hidden size降低到256再试试

name	acc	auc	f1	precision	recall	specificity
Experiment_60 fold_1 repeat	0.75	0.766666	0.636363	0.699999	0.583333	0.85
Experiment_61 fold_1 interpolate	0.78125	0.829166	0.666666	0.777777	0.583333	0.899999
Experiment_62 fold_1 random_repeat	0.8125	0.8083	0.7692	0.714285	0.833333	0.8
Experiment_63 fold_1 cyclic	0.78125	0.716666	0.72	0.6923	0.75	0.8
Experiment_64 fold_1 reflect	0.78125	0.795833	0.666666	0.777777	0.583333	0.899999
Experiment_65 fold_1 noise	0.8125	0.795833	0.75	0.75	0.75	0.85
Experiment_66 fold_1 random	0.78125	0.7875	0.695652	0.727272	0.666666	0.85

cross-ssm

name	acc	auc	f1	precision	recall	specificity
Experiment_45 fold_1	0.84375	0.858333	0.761904	0.888888	0.666666	0.949999
Experiment_45 fold_1	0.78125	0.8	0.72	0.692307	0.75	0.8
Experiment_46 fold_1	0.71875	0.704166	0.666666	0.6	0.75	0.699999
Experiment_47 fold_1	0.78125	0.758333	0.695652	0.727272	0.666666	0.85
Experiment_48 fold_1 b8	0.78125	0.779166	0.72	0.692307	0.75	0.8
Experiment_49 fold_1 b16	0.78125	0.758333	0.740740	0.666666	0.833333	0.75
增加模型feat[32, 64, 128, 256]
Experiment_50 fold_1 b8	0.78125	0.741666	0.740740	0.666666	0.833333	0.75
Experiment_51 fold_1 b8	0.78125	0.7875	0.758620	0.647058	0.916666	0.699999
Experiment_52 fold_1 b8	0.75	0.804166	0.714285	0.625	0.833333	0.699999
改回了填充到32帧
Experiment_53 fold_1	0.78125	0.779166	0.72	0.6923	0.75	0.8
Experiment_54 fold_1	0.78125	0.7625	0.666666	0.7777	0.583333	0.899999
增大hidden size 512
Experiment_56 fold_1	0.71875	0.666666	0.666666	0.6	0.75	0.699999
Experiment_57 fold_1	0.78125	0.729166	0.666666	0.777777	0.583333	0.899999
hidden size 256
Experiment_58 fold_1	0.78125	0.754166	0.740740	0.666666	0.833333	0.75
hidden size 768
Experiment_59 fold_1	0.78125	0.7875	0.6315	0.8571	0.5	0.9499

前面的损失都是bce

损失

name	acc	auc	f1	precision	recall	specificity
Experiment_67 fold_1 focal	0.6875	0.770833	0.545454	0.6	0.5	0.8
Experiment_68 fold_1 dice	0.75	0.779166	0.6	0.75	0.5	0.899999
Experiment_69 fold_1 bce_dice	0.75	0.729166	0.636363	0.699999	0.583333	0.85
Experiment_70 fold_1 tversky	0.875	0.829166	0.846153	0.785714	0.916666	0.85
Experiment_71 fold_1 asymmetric	0.78125	0.766666	0.72	0.6923	0.75	0.8

tversky 完整实验

name	acc	auc	f1	precision	recall	specificity
Experiment_70 fold_1	0.875	0.829166	0.846153	0.785714	0.916666	0.85
Experiment_70 fold_2	0.75	0.675	0.666666	0.666666	0.666666	0.8
Experiment_70 fold_3	0.71875	0.691666	0.526315	0.714285	0.416666	0.899999
Experiment_70 fold_4	0.71875	0.718614	0.608695	0.583333	0.636363	0.761904
Experiment_70 fold_5	0.5625	0.597402	0.533333	0.421052	0.727272	0.476190

还是改为16帧吧

name	acc	auc	f1	precision	recall	specificity
Experiment_74 fold_1	0.84375	0.770833	0.782608	0.818181	0.75	0.899999
Experiment_74 fold_2	0.6875	0.645833	0.375	0.75	0.25	0.949999
Experiment_74 fold_3	0.6875	0.6875	0.5454	0.6	0.5	0.8
Experiment_74 fold_4	0.78125	0.744588	0.666666	0.699999	0.636363	0.857142
Experiment_74 fold_5	0.6875	0.670995	0.583333	0.538461	0.636363	0.714285

bce损失

name	acc	auc	f1	precision	recall	specificity
Experiment_75 fold_1	0.8125	0.7666	0.75	0.75	0.75	0.85
Experiment_75 fold_2	0.6875	0.637499	0.444444	0.666666	0.333333	0.899999
Experiment_75 fold_3	0.6875	0.612499	0.583333	0.583333	0.583333	0.75
Experiment_75 fold_4	0.75	0.766233	0.555555	0.714285	0.454545	0.904761
Experiment_75 fold_5	0.59375	0.683982	0.580645	0.449999	0.818181	0.476190

去掉模块

bce

name	acc	auc	f1	precision	recall	specificity
Experiment_76 fold_1	0.8125	0.8292	0.7273	0.8	0.6667	0.8999
Experiment_76 fold_2	0.75	0.7083	0.6923	0.6429	0.75	0.75
Experiment_76 fold_3	0.75	0.7916	0.6364	0.6999	0.5833	0.85
Experiment_76 fold_4	0.78125	0.7359	0.6666	0.6999	0.6363	0.8571
Experiment_76 fold_5	0.4375	0.5757	0.5263	0.3703	0.9090	0.1904

只加上gated

name	acc	auc	f1	precision	recall	specificity
Experiment_77 fold_1	0.6875	0.7375	0.6666	0.5555	0.8333	0.6
Experiment_77 fold_2	0.7187	0.6666	0.64	0.6153	0.6666	0.75
Experiment_77 fold_3	0.5937	0.5916	0.6285	0.4782	0.9166	0.4
Experiment_77 fold_4	0.5625	0.6147	0.5882	0.4347	0.9090	0.3809
Experiment_77 fold_5

只加上attn

name	acc	auc	f1	precision	recall	specificity
Experiment_78 fold_1	0.75	0.7541	0.6666	0.6666	0.6666	0.8
Experiment_78 fold_2	0.71875	0.6916	0.64	0.6153	0.6666	0.75
Experiment_78 fold_3	0.625	0.6208	0.6	0.5	0.75	0.55
Experiment_78 fold_4	0.7187	0.5930	0.5263	0.625	0.4545	0.8571
Experiment_78 fold_5

3.1.1 mamba+echoprime video

使用两个分支对两个切面的数据进行提取，echo_prime和mamba进行提取，最后一起送入分类器

mamba: A2C切面的数据

echo_prime: A4C切面的数据

融合: 通道拼接

name	acc	auc	f1	precision	recall	specificity
Experiment_79 fold_1	0.7813	0.7333	0.6667	0.7777	0.5833	0.90
Experiment_79 fold_2	0.78125	0.774999	0.5882	1.0	0.4166	1.0
Experiment_79 fold_3	0.875	0.9208	0.818181	0.899999	0.75	0.949999
Experiment_79 fold_4	0.84375	0.839826	0.761904	0.8	0.727272	0.904761
Experiment_79 fold_5	0.78125	0.757575	0.6666	0.699999	0.636363	0.857142
Experiment_79	0.8125

保留2*2*2的空间结构

name	acc	auc	f1	precision	recall	specificity
Experiment_80 fold_1	0.78125	0.695833	0.666666	0.777777	0.5833	0.8999
Experiment_80 fold_2	0.78125	0.787499	0.695652	0.727272	0.666666	0.85
Experiment_80 fold_3	0.875	0.8958	0.846153	0.785714	0.916666	0.85
Experiment_80 fold_4	0.84375	0.839826	0.782608	0.75	0.818181	0.857142
Experiment_80 fold_5	0.8125	0.787878	0.6666	0.8571	0.5454	0.9523
Experiment_80	0.8188	0.8014	0.7315	0.7796	0.7060	0.8818

保留4*4*4的空间结构

name	acc	auc	f1	precision	recall	specificity
Experiment_81 fold_1	0.8125	0.875	0.6999	0.875	0.5833	0.9499
Experiment_81 fold_2	0.8125	0.775	0.7272	0.8	0.666666	0.899999
Experiment_81 fold_3	0.875	0.879166	0.8333	0.8333	0.8333	0.8999
Experiment_81 fold_4	0.84375	0.848484	0.7826	0.75	0.818181	0.857142
Experiment_81 fold_5	0.84375	0.7922	0.7619	0.8	0.727272	0.904761
Experiment_80	0.8375	0.8139	0.7610	0.8117	0.7257	0.9023

交换特征提取器

mamba: A4C切面的数据

echo_prime: A2C切面的数据

融合: 通道拼接

保留4*4*4的空间结构

name	acc	auc	f1	precision	recall	specificity
Experiment_82 fold_1	0.90625	0.879166	0.879999	0.846153	0.916666	0.899999
Experiment_82 fold_2	0.78125	0.741666	0.695652	0.727272	0.666666	0.85
Experiment_82 fold_3	0.8125	0.8666	0.7692	0.714285	0.8333	0.8
Experiment_82 fold_4	0.90625	0.9220	0.842105	1.0	0.727272	1.0
Experiment_82 fold_5	0.75	0.6666	0.5	0.8	0.363636	0.9523
Experiment_82	0.8312	0.8152	0.7373	0.8175	0.6976	0.8805

学习率

name	acc	auc	f1	precision	recall	specificity
Experiment_83 fold_5 1e-4	0.71875	0.6320	0.3077	1.0	0.1818	1.0
Experiment_84 fold_5 1e-6	0.78125	0.7445	0.5882	0.8333	0.454545	0.952380

name	acc	auc	f1	precision	recall	specificity
Experiment_84 fold_1	0.9375	0.904166	0.916666	0.916666	0.916666	0.949999
Experiment_84 fold_2	0.78125	0.699999	0.72	0.692307	0.75	0.8
Experiment_84 fold_3	0.84375	0.816666	0.8	0.769230	0.833333	0.85
Experiment_84 fold_4	0.90625	0.8961	0.8571	0.8999	0.818181	0.952380
Experiment_84 fold_5	0.78125	0.7445	0.5882	0.8333	0.454545	0.952380
Experiment_84 95%	0.8500	0.8249	0.7857	0.8148	0.7586	0.9020

3.1.2 mamba+echoprime text video

在mamba的decoder部分加入三层知识向量

mamba: A4C切面的数据

echo_prime: A2C切面的数据

融合: 通道拼接

保留4*4*4的空间结构 损失bce

name	acc	auc	f1	precision	recall	specificity
Experiment_102 fold_1	0.9375	0.895833	0.916666	0.916666	0.916666	0.949999
Experiment_102 fold_2	0.78125	0.7374	0.6666	0.7777	0.5833	0.8999
Experiment_102 fold_3	0.78125
Experiment_102 fold_4	0.875
Experiment_102 fold_5	0.75
Experiment_102	0.816

在mamba的decoder部分加入知识向量和echo prime视频提取部分都加入知识向量

学习率1e-5

name	acc	auc	f1	precision	recall	specificity
Experiment_103 fold_1	0.90625	0.8791	0.8799	0.8461	0.9166	0.8999
Experiment_103 fold_2	0.78125
Experiment_103 fold_3	0.8125
Experiment_103 fold_4	0.875	0.8744	0.8	0.8888	0.7272	0.9523
Experiment_103 fold_5	0.75
Experiment_103	0.825

损失asymmetric 学习率1e-5

name	acc	auc	f1	precision	recall	specificity
Experiment_104 fold_1	0.90625
Experiment_104 fold_2	0.78125
Experiment_104 fold_5	0.75

学习率

name	acc	auc	f1	precision	recall	specificity
Experiment_105 fold_5 1e-6	0.71875
Experiment_106 fold_5 1e-5	0.75
Experiment_107 fold_5 1e-4	0.71875
Experiment_108 fold_5 3e-4	0.75
Experiment_109 fold_5 3e-5	0.6875
Experiment_110 fold_5 3e-6	0.71875

效果好差，还是只加一层知识向量吧

name	acc	auc	f1	precision	recall	specificity
Experiment_111 fold_5 1e-6	0.75

在A4C的mamba的瓶颈层加入知识向量，同时A2C的video encoder也加入知识向量

name	acc	auc	f1	precision	recall	specificity
Experiment_112 fold_1	0.9375	0.925	0.916666	0.916666	0.916666	0.9499
Experiment_112 fold_2	0.78125	0.7374	0.6666	0.7777	0.5833	0.8999
Experiment_112 fold_3	0.78125	0.7958	0.72	0.6923	0.75	0.8
Experiment_112 fold_4	0.875	0.8701	0.8181	0.818181	0.818181	0.9047
Experiment_112 fold_5
Experiment_112

b16

name	acc	auc	f1	precision	recall	specificity
Experiment_113 fold_1	0.90625	0.9208	0.8799	0.8461	0.9166	0.8999
Experiment_113 fold_2	0.78125	0.725	0.72	0.6923	0.75	0.8
Experiment_113 fold_3	0.75	0.8125	0.7142	0.625	0.8333	0.6999
Experiment_113 fold_4	0.90625	0.8658	0.8571	0.8999	0.8181	0.9523
Experiment_113 fold_5	0.78125	0.6666	0.6666	0.6999	0.6363	0.8571
Experiment_113

交换A2C和A4C 十分拉跨

name	acc	auc	f1	precision	recall	specificity
Experiment_114 fold_1	0.75	0.8458	0.7333	0.6111	0.9166	0.6499
Experiment_114 fold_2	0.8125	0.875	0.7692	0.7142	0.8333	0.8
Experiment_114 fold_3	0.78125	0.825	0.7586	0.6470	0.9166	0.6999
Experiment_114 fold_4	0.6875	0.7748	0.6428	0.5294	0.8181	0.6190
Experiment_114 fold_5	0.75	0.6926	0.6666	0.6153	0.7272	0.7619

在mamba的decoder部分加入知识向量和echo prime video encoder 加入知识向量

mamba: A4C切面的数据

echo_prime: A2C切面的数据

融合: 通道拼接

保留4*4*4的空间结构 损失bce

学习率1e-6 b8

name	acc	auc	f1	precision	recall	specificity
Experiment_115 fold_1	0.9375	0.9125	0.9166	0.9166	0.9166	0.9499
Experiment_115 fold_2	0.8125	0.7333	0.75	0.75	0.75	0.85
Experiment_115 fold_3	0.8125	0.8041	0.7272	0.8	0.6666	0.8999
Experiment_115 fold_4	0.90625	0.9004	0.8571	0.8999	0.8181	0.9523
Experiment_115 fold_5	0.78125	0.7835	0.5882	0.8333	0.4545	0.9523
Experiment_115 95%	0.8500	0.8315	0.7778	0.8400	0.7241	0.9216

3.1.3 orginmutimodel

name	acc	auc	f1	precision	recall	specificity
Experiment_132 fold_1	0.78125	0.7041	0.6315	0.8571	0.5	0.9499
Experiment_132 fold_2	0.78125	0.7083	0.6666	0.7777	0.5833	0.8999
Experiment_132 fold_3	0.71875	0.6666	0.64	0.6153	0.6666	0.75
Experiment_132 fold_4	0.8125	0.8095	0.75	0.6923	0.8181	0.8095
Experiment_132 fold_5	0.6562	0.4155	0.0	0.0	0.0	1.0
Experiment_132

3.2 hmc数据集

简单的先尝试，效果非常的糟糕

name	acc	auc	f1	precision	recall	specificity
Experiment_73 fold_1	0.84375	0.884057	0.888888	0.909090	0.869565	0.777777
Experiment_73 fold_2	0.75	0.744588	0.826086	0.759999	0.904761	0.454545
Experiment_73 fold_3	0.6875	0.647058	0.75	0.652174	0.882352	0.466666
Experiment_73 fold_4	0.6875	0.536796	0.807692	0.677419	1.0	0.090909
Experiment_73 fold_5

3.2.1 mamba+echoprime video

mamba: A4C切面的数据

echo_prime: A2C切面的数据

融合: 通道拼接

保留4*4*4的空间结构

name	acc	auc	f1	precision	recall	specificity
Experiment_85 fold_1	0.90625	0.913043	0.936170	0.916666	0.956521	0.777777
Experiment_85 fold_2	0.78125	0.787878	0.829268	0.85	0.8095	0.7272
Experiment_85 fold_3	0.8125	0.8784	0.8235	0.8235	0.8235	0.8
Experiment_85 fold_4	0.8125	0.8095	0.85	0.8947	0.8095	0.8182
Experiment_85 fold_5	0.84375	0.878787	0.87179	0.9444	0.8095	0.9090
Experiment_85	0.83125

损失改为tversky

name	acc	auc	f1	precision	recall	specificity
Experiment_86 fold_2	0.78125	0.813852	0.820512	0.888888	0.761904	0.818181

损失改为focal

name	acc	auc	f1	precision	recall	specificity
Experiment_87 fold_2	0.78125	0.8	0.8108	0.9375	0.714285	0.9090

损失改为dice

name	acc	auc	f1	precision	recall	specificity
Experiment_88 fold_2	0.78125	0.783549	0.820512	0.888888	0.761904	0.818181

损失改为bce_dice

name	acc	auc	f1	precision	recall	specificity
Experiment_89 fold_2	0.78125	0.805194	0.8205	0.8888	0.7619	0.818181

损失改为asymmetric

name	acc	auc	f1	precision	recall	specificity
Experiment_90 fold_1	0.90625	0.932367	0.936170	0.916666	0.956521	0.777777
Experiment_90 fold_2	0.8125	0.8484	0.8571	0.8571	0.8571	0.7272
Experiment_90 fold_3	0.84375	0.843137	0.864864	0.8	0.941176	0.733333
Experiment_90 fold_4	0.8125	0.835497	0.85	0.894736	0.809523	0.818181
Experiment_90 fold_5	0.84375	0.861471	0.883720	0.863636	0.904761	0.727272
Experiment_90	0.84375	0.864175	0.888351	0.866628	0.891956	0.753332

改为交叉注意力融合

name	acc	auc	f1	precision	recall	specificity
Experiment_95 fold_1	0.84375	0.903381	0.883720	0.949999	0.826086	0.888888
Experiment_95 fold_2	0.78125	0.779220	0.851063	0.769230	0.952380	0.454545
Experiment_95 fold_3	0.8125
Experiment_95 fold_4	0.78125
Experiment_95 fold_5	0.8125	0.818181	0.857142	0.857142	0.857142	0.727272
Experiment_95	拉跨

加权融合

name	acc	auc	f1	precision	recall	specificity
Experiment_96 fold_1	0.90625	0.937198	0.936170	0.916666	0.9565	0.7777
Experiment_96 fold_2	0.8125	0.826839	0.857142	0.857142	0.857142	0.727272
Experiment_96 fold_3	0.8125	0.882352	0.849999	0.739130	1.0	0.6
Experiment_96 fold_4	0.8125	0.826839	0.85	0.894736	0.809523	0.818181
Experiment_96 fold_5	0.84375	0.857142	0.883720	0.863636	0.904761	0.727272
Experiment_96

Adaw优化器

name	acc	auc	f1	precision	recall	specificity
Experiment_97 fold_1	0.90625	0.927536	0.930232	1.0	0.869565	1.0
Experiment_97 fold_2	0.78125	0.831168	0.810810	0.9375	0.714285	0.909090
Experiment_97 fold_3	0.78125	0.882352	0.774193	0.857142	0.705882	0.866666
Experiment_97 fold_4	0.8125	0.844155	0.842105	0.941176	0.761904	0.909090
Experiment_97 fold_5	0.8125	0.883116	0.85	0.894736	0.809523	0.818181
Experiment_97

SGD

name	acc	auc	f1	precision	recall	specificity
Experiment_98 fold_1	0.8125	0.816425	0.869565	0.869565	0.869565	0.666666
Experiment_98 fold_2	0.65625

RMSprop

name	acc	auc	f1	precision	recall	specificity
Experiment_99 fold_1	0.90625	0.913043	0.930232	1.0	0.869565	1.0
Experiment_99 fold_2	0.78125	0.822510	0.810810	0.9375	0.714285	0.909090

Adagrad

name	acc	auc	f1	precision	recall	specificity
Experiment_100 fold_1	0.78125
Experiment_100 fold_2	0.65625

还是改回MADGRAD优化器

做了点数据增强

name	acc	auc	f1	precision	recall	specificity
Experiment_101 fold_1	0.875	0.9565	0.916666	0.879999	0.956521	0.666666
Experiment_101 fold_2	0.8125	0.7922	0.85	0.8947	0.8095	0.818181
Experiment_101 fold_3	0.84375	0.886274	0.871794	0.772727	1.0	0.666666
Experiment_101 fold_4	0.75
Experiment_101 fold_5	0.8125	0.8311	0.875	0.7777	1.0	0.4545

mamba: A2C切面的数据

echo_prime: A4C切面的数据

融合: 通道拼接

保留4*4*4的空间结构

损失还是asymmetric

name	acc	auc	f1	precision	recall	specificity
Experiment_92 fold_1	0.90625	0.971014	0.936170	0.916666	0.956521	0.7777
Experiment_92 fold_2	0.84375	0.805194	0.888888	0.833333	0.952380	0.636363
Experiment_92 fold_3	0.75	0.7960	0.7647	0.7647	0.7647	0.7333
Experiment_92 fold_4	0.78125	0.826839	0.810810	0.9375	0.714285	0.909090
Experiment_92 fold_5	0.84375	0.878787	0.878048	0.899999	0.857142	0.818181
Experiment_92	0.825	效果拉跨

mamba: A4C切面的数据

echo_prime: A2C切面的数据

融合: 通道拼接

保留4*4*4的空间结构

损失还是asymmetric

但是使用的是dec0作为mamba的输出

name	acc	auc	f1	precision	recall	specificity
Experiment_93 fold_1	0.90625	0.942028	0.938775	0.884615	1.0	0.666666
Experiment_93 fold_2	0.8125	0.792207	0.85	0.894736	0.809523	0.818181
Experiment_93 fold_3	0.8125	0.870588	0.823529	0.823529	0.823529	0.8
Experiment_93 fold_4	0.8125	0.805194	0.857142	0.857142	0.857142	0.727272
Experiment_93 fold_5	0.84375	0.874458	0.878048	0.89999	0.857142	0.818181
Experiment_93	0.8375

保留6*6*6的空间结构

name	acc	auc	f1	precision	recall	specificity
Experiment_93 fold_1	0.90625
Experiment_93 fold_3	0.75

mamba: A4C切面的数据

echo_prime: A2C切面的数据

融合: 通道拼接

保留4*4*4的空间结构

损失还是asymmetric

但是使用的是dec2作为mamba的输出

name	acc	auc	f1	precision	recall	specificity
Experiment_94 fold_1	0.90625	0.946859	0.936170	0.916666	0.956521	0.777777
Experiment_94 fold_2	0.78125	0.796536	0.829268	0.85	0.809523	0.727272
Experiment_94 fold_3	0.84375	0.925490	0.848484	0.875	0.823529	0.866666
Experiment_94 fold_4	0.75	0.779220	0.809523	0.809523	0.809523	0.636363
Experiment_94 fold_5	0.8125	0.861471	0.857142	0.857142	0.857142	0.727272

name	acc	auc	f1	precision	recall	specificity
Experiment_95 fold_1 1e-8	0.90625	0.9565	0.938775	0.884615	1.0	0.666666
Experiment_95 fold_3 1e-8	0.8125	0.921568	0.849999	0.739130	1.0	0.6

3.2.2 mamba+echoprime text video

加入知识向量

mamba: A4C切面的数据

echo_prime: A2C切面的数据

都加入了知识向量

融合: 通道拼接

保留4*4*4的空间结构

学习率1e-6 b8 损失：bce

name	acc	auc	f1	precision	recall	specificity
Experiment_116 fold_1	0.90625	0.8067	0.9387	0.8846	1.0	0.6666
Experiment_116 fold_2	0.78125 (0.81)	0.818181	0.8108	0.9375	0.7142	0.9090
Experiment_116 fold_3	0.8125 (0.84)	0.8509	0.8	0.9230	0.7058	0.9333
Experiment_116 fold_4	0.8125	0.8614	0.85	0.8947	0.8095	0.8181
Experiment_116 fold_5	0.84375	0.8484	0.8837	0.8636	0.9047	0.7272
Experiment_116	0.83125

name	acc	auc	f1	precision	recall	specificity
Experiment_127 fold_1	0.90625	0.8792	0.9387	0.8846	1.0	0.6666
Experiment_127 fold_2	0.8125	0.9264	0.8333	1.0	0.7142	1.0
Experiment_127 fold_3	0.8125	0.8666	0.8	0.9230	0.7058	0.9333
Experiment_127 fold_4	0.75	0.8138	0.7894	0.8823	0.7142	0.8181
Experiment_127 fold_5	0.8125	0.8225	0.8571	0.8571	0.8571	0.7272
Experiment_127

学习率1e-6 b16 损失：bce

name	acc	auc	f1	precision	recall	specificity
Experiment_128 fold_1	0.875	0.9178	0.9166	0.8799	0.9565	0.6666
Experiment_128 fold_2	0.75	0.7878	0.7894	0.8823	0.7142	0.8181
Experiment_128 fold_3	0.84375	0.9098	0.8387	0.9285	0.7647	0.9333
Experiment_128 fold_4	0.78125	0.8225	0.8108	0.9375	0.7142	0.9090
Experiment_128 fold_5	0.84375	0.8658	0.8780	0.8999	0.8571	0.8181
Experiment_128	0.81875	0.86072	0.84474	0.90562	0.78136	0.84722

学习率1e-6 b12 损失：bce

name	acc	auc	f1	precision	recall	specificity
Experiment_129 fold_1	0.90625	0.9371	0.9361	0.9166	0.9565	0.7777
Experiment_129 fold_2	0.75	0.7922	0.7999	0.8421	0.7619	0.7272
Experiment_129 fold_3	0.84375	0.8549	0.8387	0.9285	0.7647	0.9333
Experiment_129 fold_4	0.75	0.7965	0.7894	0.8823	0.7142	0.8181
Experiment_129 fold_5	0.84375	0.8614	0.8717	0.9444	0.8095	0.9090
Experiment_129

学习率1e-6 b4 损失：bce

name	acc	auc	f1	precision	recall	specificity
Experiment_130 fold_1	0.90625	0.9516	0.9333	0.9545	0.9130	0.8888
Experiment_130 fold_2	0.78125	0.7835	0.8372	0.8181	0.8571	0.6363
Experiment_130 fold_3	0.84375	0.8784	0.8387	0.9285	0.7647	0.9333
Experiment_130 fold_4	0.78125	0.8	0.8205	0.8888	0.7619	0.8181
Experiment_130 fold_5	0.84375	0.8484	0.8717	0.9444	0.8095	0.9090
Experiment_130

name	acc	auc	f1	precision	recall	specificity
Experiment_131 fold_1	0.9375	0.9565	0.9545	1.0	0.9130	1.0
Experiment_131 fold_2	0.8125	0.8701	0.85	0.8947	0.8095	0.8181
Experiment_131 fold_3	0.875	0.9568	0.8947	0.8095	1.0	0.9333
Experiment_131 fold_4	0.8125	0.8744	0.85	0.8947	0.8095	0.8181
Experiment_131 fold_5	0.875	0.8961	0.8999	0.9473	0.8571	0.9090
Experiment_131	0.8625	0.9108	0.8899	0.9093	0.8778	0.8558

name	acc	auc	f1	precision	recall	specificity
Experiment_131 fold_5 v1	0.84375	0.8831	0.8717	0.9444	0.8095	0.9090
Experiment_131 fold_5 v2	0.84375	0.9047	0.8780	0.8999	0.8571	0.8181

使用dec2作为输出

name	acc	auc	f1	precision	recall	specificity
Experiment_126 fold_1	0.90625	0.8937	0.9387	0.8846	1.0	0.6666
Experiment_126 fold_2	0.8125	0.8571	0.8571	0.8571	0.8571	0.7272
Experiment_126 fold_3	0.75	0.8117	0.75	0.8	0.7058	0.8
Experiment_126 fold_4	0.6875	0.7662	0.75	0.7894	0.7142	0.6363
Experiment_126 fold_5	0.84375	0.8354	0.8718	0.9444	0.8095	0.9090
Experiment_126

学习率1e-5 b8 损失：bce

name	acc	auc	f1	precision	recall	specificity
Experiment_125 fold_1	0.875	0.8743	0.9166	0.8799	0.9565	0.6666
Experiment_125 fold_2	0.75	0.8051	0.7777	0.9333	0.6666	0.9090
Experiment_125 fold_3	0.78125	0.8705	0.7407	1.0	0.5882	1.0
Experiment_125 fold_4	0.71875	0.7445	0.7567	0.875	0.6666	0.8181
Experiment_125 fold_5	0.84375	0.8311	0.8837	0.8636	0.9047	0.7272
Experiment_125

损失：asymmetric

name	acc	auc	f1	precision	recall	specificity
Experiment_117 fold_1	0.90625	0.8405	0.9387	0.8846	1.0	0.6666
Experiment_117 fold_2	0.75	0.7229	0.7777	0.9333	0.6666	0.9090
Experiment_117 fold_3	0.8125 (0.84)	0.8823	0.8421	0.7619	0.9411	0.6666
Experiment_117 fold_4	0.8125	0.8528	0.8421	0.9411	0.7619	0.9090
Experiment_117 fold_5	0.84375	0.8354	0.8837	0.8636	0.9047	0.7272
Experiment_117	0.825

损失weighted_bce 2.0

name	acc	auc	f1	precision	recall	specificity
Experiment_118 fold_1	0.875	0.8937	0.9166	0.8799	0.9565	0.6666
Experiment_118 fold_2	0.78125	0.818181	0.8205	0.8888	0.7619	0.8181
Experiment_118 fold_3	0.8125	0.8431	0.8	0.9230	0.7058	0.9333
Experiment_118 fold_4	0.75	0.8354	0.7894	0.8823	0.7142	0.8181
Experiment_118 fold_5	0.84375	0.8441	0.8837	0.8636	0.9047	0.7272
Experiment_118	0.8125	0.8469	0.8420	0.8875	0.7886	0.7931

损失weighted_bce 0.75

name	acc	auc	f1	precision	recall	specificity
Experiment_119 fold_1	0.90625	0.8550	0.9387	0.8846	1.0	0.6666
Experiment_119 fold_2	0.78125	0.8138	0.8108	0.9375	0.7142	0.9090
Experiment_119 fold_3	0.84375	0.9098	0.8571	0.8333	0.8823	0.8
Experiment_119 fold_4	0.8125	0.8528	0.8571	0.8571	0.8571	0.7272
Experiment_119 fold_5	0.8125	0.8528	0.8571	0.8571	0.8571	0.7272
Experiment_119	0,83125

损失weighted_bce 0.5

name	acc	auc	f1	precision	recall	specificity
Experiment_120 fold_1	0.84375	0.8550	0.8936	0.875	0.9130	0.6666
Experiment_120 fold_2	0.75	0.8268	0.7777	0.9333	0.6666	0.9090
Experiment_120 fold_3	0.875	0.9058	0.8888	0.8421	0.9411	0.8
Experiment_120 fold_4	0.75	0.8398	0.7894	0.8823	0.7142	0.8181
Experiment_120 fold_5	0.78125	0.8398	0.8292	0.85	0.8095	0.7272
Experiment_120	0.8

使用的dec0作为mamba的输出

学习率1e-6 b8 损失：bce

name	acc	auc	f1	precision	recall	specificity
Experiment_121 fold_1	0.875	0.9227	0.9166	0.8799	0.9565	0.6666
Experiment_121 fold_2	0.8125	0.8051	0.875	0.7777	1.0	0.4545
Experiment_121 fold_3	0.875	0.8862	0.8823	0.8823	0.8823	0.8666
Experiment_121 fold_4	0.6875	0.6839	0.7222	0.8666	0.6190	0.8181
Experiment_121 fold_5	0.875	0.8917	0.9047	0.9047	0.9047	0.8181
Experiment_121	0.825

使用的dec0作为mamba的输出

学习率1e-6 b8 损失：asymmetric

name	acc	auc	f1	precision	recall	specificity
Experiment_122 fold_4	0.6875

学习率:1e-4

name	acc	auc	f1	precision	recall	specificity
Experiment_123 fold_1	0.875	0.8550	0.9166	0.8799	0.9565	0.6666
Experiment_123 fold_2	0.75
Experiment_123 fold_3	0.84375
Experiment_123 fold_4	0.71875	0.7532	0.7428	0.9285	0.6190	0.9090
Experiment_123 fold_5	0.8125
Experiment_123

学习率1e-4 b8 损失：bce 少加几层知识向量

name	acc	auc	f1	precision	recall	specificity
Experiment_124 fold_1	0.90625	0.8599	0.9387	0.8846	1.0	0.6666
Experiment_124 fold_2	0.78125	0.8	0.8205	0.8888	0.7619	0.8181
Experiment_124 fold_3	0.84375	0.9568	0.8387	0.9285	0.7647	0.9333
Experiment_124 fold_4	0.71875	0.7532	0.7428	0.9285	0.6190	0.9090
Experiment_124 fold_5	0.8125	0.8311	0.8571	0.8571	0.8571	0.7272
Experiment_124	0.8125

3.3 省医院数据

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_test fold_5

A2C A4C

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_test fold_1	0.75	0.7219	0.8108	0.75	0.8823	0.5454
Experiment_provincial_test fold_2	0.75	0.8556	0.8205	0.7272	0.9411	0.4545
Experiment_provincial_test fold_3	0.8214	0.8214	0.8484	0.7368	1.0	0.6428
Experiment_provincial_test fold_4	0.6666	0.4671	0.7804	0.7272	0.8421	0.25
Experiment_provincial_test fold_5	0.7407	0.6882	0.7878	0.8125	0.7647	0.6999
Experiment_provincial_test	0.74

6切面的数据 mamba

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_1 fold_1	0.8571	0.8877	0.8823	0.8823	0.8823	0.8181
Experiment_provincial_1 fold_2	0.8571	0.9197	0.875	0.9333	0.8235	0.9090
Experiment_provincial_1 fold_3	0.8214	0.8775	0.8275	0.8	0.8571	0.7857
Experiment_provincial_1 fold_4	0.7692	0.5486	0.8571	0.75	1.0	0.25
Experiment_provincial_1 fold_5	0.7307	0.6405	0.7741	0.8571	0.7058	0.7777
Experiment_provincial_1	0.8088	0.7881	0.8452	0.8353	0.8554	0.7358

6切面的数据 mamba 三分类

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_3 fold_1	0.75	0.6893	0.75	0.75	0.75
Experiment_provincial_3 fold_2	0.7142	0.8654	0.7142	0.7142	0.7142
Experiment_provincial_3 fold_3	0.75	0.8795	0.75	0.75	0.75
Experiment_provincial_3 fold_4	0.8076	0.6731	0.8076	0.8076	0.8076
Experiment_provincial_3 fold_5	0.7307	0.6515	0.7307	0.7307	0.7307
Experiment_provincial_3	0.7505	0.7518	0.7505	0.7505	0.7505

430例

name	acc	auc	f1	precision	recall
Experiment_provincial_15 fold_0	0.5909	0.6442	0.5909	0.5909	0.5909
Experiment_provincial_15 fold_1	0.6351	0.5633	0.6351	0.6351	0.6351
Experiment_provincial_15 fold_2	0.6279	0.4458	0.6279	0.6279	0.6279
Experiment_provincial_15 fold_3	0.6304	0.4827	0.6304	0.6304	0.6304
Experiment_provincial_15 fold_4	0.5595	0.5470	0.5595	0.5595	0.5595
Experiment_provincial_15

name	acc	auc	f1	precision	recall
Experiment_provincial_16 fold_0	0.6022	0.6275	0.6022	0.6022	0.6022
Experiment_provincial_16 fold_1	0.6351	0.5748	0.6351	0.6351	0.6351
Experiment_provincial_16 fold_2	0.6395	0.5086	0.6395	0.6395	0.6395
Experiment_provincial_16 fold_3	0.6413	0.4055	0.6413	0.6413	0.6413
Experiment_provincial_16 fold_4	0.6071	0.6992	0.6071	0.6071	0.6071
Experiment_provincial_16	0.6250	0.5631	0.6250	0.6250	0.6250

6切面的数据 echoprime 32F

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_2 fold_1	0.8125	0.9047	0.8571	0.75	1.0	0.5714
Experiment_provincial_2 fold_2	0.8125	0.6727	0.8799	0.7857	1.0	0.4
Experiment_provincial_2 fold_3	0.8125	0.8166	0.8	0.6666	1.0	0.6999
Experiment_provincial_2 fold_4	0.8125	0.5	0.8888	0.8	1.0	0.25
Experiment_provincial_2 fold_5	0.625	0.55	0.7692	0.625	1.0	0.0
Experiment_provincial_2

加上12月的数据

六切面 16F

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_4 fold_1	0.75	0.7366	0.8346	0.7361	0.9636	0.3448
Experiment_provincial_4 fold_2	0.7093	0.7020	0.7788	0.7097	0.8627	0.4857
Experiment_provincial_4 fold_3	0.7262	0.7312	0.7473	0.7234	0.7727	0.6750
Experiment_provincial_4 fold_4	0.6860	0.6577	0.7652	0.6769	0.8799	0.4166
Experiment_provincial_4 fold_5	0.7143	0.6695	0.7931	0.7302	0.8679	0.4516
Experiment_provincial_4

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_13 fold_1	0.7857	0.8012	0.8448	0.8032	0.8909	0.5862
Experiment_provincial_13 fold_2	0.6744	0.6224	0.7666	0.6666	0.9020	0.3428
Experiment_provincial_13 fold_3	0.75	0.7540	0.7470	0.7949	0.7045	0.8
Experiment_provincial_13 fold_4	0.6860	0.6277	0.7768	0.6620	0.9399	0.3333
Experiment_provincial_13 fold_5	0.6904	0.6524	0.7592	0.7454	0.7736	0.5484
Experiment_provincial_13	0.7173

新分折

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_14 fold_0	0.6931	0.6963	0.7216	0.7609	0.6862	0.7027
Experiment_provincial_14 fold_1	0.7567	0.7720	0.7954	0.7954	0.7954	0.6999
Experiment_provincial_14 fold_2	0.7209	0.7020	0.7857	0.7586	0.8148	0.5625
Experiment_provincial_14 fold_3	0.6739	0.7306	0.6875	0.8461	0.5789	0.8286
Experiment_provincial_14 fold_4	0.5952	0.5399	0.6909	0.6129	0.7917	0.3333
Experiment_provincial_14	0.6880

A2C + A4C

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_9 fold_1	0.8	0.8026	0.8491	0.8181	0.8824	0.6552
Experiment_provincial_9 fold_2	0.7250	0.7104	0.7755	0.7169	0.8444	0.5714
Experiment_provincial_9 fold_3	0.75	0.7654	0.7727	0.7234	0.8293	0.6666
Experiment_provincial_9 fold_4	0.75	0.7644	0.7561	0.8378	0.6888	0.8286
Experiment_provincial_9 fold_5	0.7125	0.6466	0.7965	0.7143	0.8999	0.4
Experiment_provincial_9	0.7475	0.7379	0.79	0.7621	0.8290	0.6244

新分折

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_10 fold_0	0.6704
Experiment_provincial_10 fold_1	0.6527
Experiment_provincial_10 fold_2	0.725
Experiment_provincial_10 fold_3
Experiment_provincial_10 fold_4
Experiment_provincial_10

A2C + A3C + A4C

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_11 fold_1	0.8095	0.8508	0.8545	0.8545	0.8545	0.7241
Experiment_provincial_11 fold_2	0.7381	0.7038	0.7924	0.7368	0.8571	0.5714
Experiment_provincial_11 fold_3	0.75	0.7897	0.7640	0.7555	0.7727	0.725
Experiment_provincial_11 fold_4	0.75	0.7481	0.7921	0.7692	0.8163	0.6571
Experiment_provincial_11 fold_5	0.7381	0.6962	0.8	0.7719	0.8301	0.5806
Experiment_provincial_11	0.7571	0.7577	0.8006	0.7776	0.8262	0.6517

AP SAX + MV SAX + PM SAX

name	acc	auc	f1	precision	recall	specificity
Experiment_provincial_12 fold_1	0.7976	0.8132	0.8468	0.8393	0.8545	0.6896
Experiment_provincial_12 fold_2	0.7619	0.7738	0.8	0.7843	0.8163	0.6857
Experiment_provincial_12 fold_3	0.7857	0.8159	0.7857	0.8250	0.75	0.8250
Experiment_provincial_12 fold_4	0.7619	0.7796	0.8113	0.7544	0.8775	0.60
Experiment_provincial_12 fold_5	0.7381	0.7553	0.7924	0.7924	0.7924	0.6451
Experiment_provincial_12	0.7690	0.7876	0.8073	0.7991	0.8182	0.6891

4. 模型结构图

5. sota实验

5.1 camus

汇总

name	acc	auc	f1	precision	recall	specificity
CARL	83.75	60.33	73.42	85.56	67.42	/
MV-Swin-T	62.50	57.14	60.42	61.65	62.17	/
CTT-Net	62.50	70.04	70.70	62.96	82.62	/
E-ViM³
E-ViM³ + plugin
XFMamba	71.88	57.08	47.05	80.00	33.33
XFMamba + plugin	76.87	77.87	65.04	73.02	60.00
BI-Mamba	68.75	50.21	37.50	60.00	27.27
BI-Mamba + plugin	76.88	76.68	58.94	81.05	51.66
DK-Mamba	85.00	83.15	77.78	84.00	72.41	92.16

BI-mamba

name	acc	auc	f1	precision	recall
Experiment_158 fold_1	0.6875	0.3541	0.375	0.75	0.25
Experiment_158 fold_2	0.6562	0.4250	0.1538	1.0	0.0833
Experiment_158 fold_3	0.625	0.4249	0	0	0
Experiment_158 fold_4	0.6875	0.5021	0.3750	0.6000	0.2727
Experiment_158 fold_5	0.65625	0.5064	0	0	0
Experiment_158

BI-mamba + plugin

name	acc	auc	f1	precision	recall
Experiment_157 fold_1	0.875	0.8541	0.8333	0.8333	0.8333
Experiment_157 fold_2	0.6875	0.6666	0.2857	1.0	0.1666
Experiment_157 fold_3	0.71875	0.7333	0.6087	0.6363	0.5833
Experiment_157 fold_4	0.78125	0.8138	0.5882	0.8333	0.4545
Experiment_157 fold_5	0.78125	0.7662	0.6315	0.75	0.5454
Experiment_157	0.7688	0.7668	0.5894	0.8105	0.5166

xf-mamba

name	acc	auc	f1	precision	recall
Experiment_159 fold_1	0.65625	0.2791	0.2666	0.6666	0.1666
Experiment_159 fold_2	0.625	0.5499	0	0	0
Experiment_159 fold_3	0.71875	0.5708	0.4705	0.8000	0.3333
Experiment_159 fold_4	0.6875	0.6060	0.1666	1.0	0.09
Experiment_159 fold_5	0.65625	0.6709	0.1538	0.5	0.09
Experiment_159

xf-mamba-plugin

name	acc	auc	f1	precision	recall
Experiment_160 fold_1	0.8437	0.8541	0.7826	0.8181	0.75
Experiment_160 fold_2	0.6875	0.7041	0.5833	0.5833	0.5833
Experiment_160 fold_3	0.75	0.7291	0.6666	0.6666	0.6666
Experiment_160 fold_4	0.78125	0.8225	0.5882	0.8333	0.4545
Experiment_160 fold_5	0.7812	0.7835	0.6315	0.75	0.5454
Experiment_160	0.76873	0.77866	0.65044	0.73026	0.59996

E-ViM³

name	acc	auc	f1	precision	recall
Experiment_161 fold_1
Experiment_161 fold_2
Experiment_161 fold_3
Experiment_161 fold_4
Experiment_161 fold_5
Experiment_161

5.2 hmc

汇总

name	acc	auc	f1	precision	recall	specificity
SAF-Net	78.13	\	81.57	88.26	77.64	/
CARL	85.62	45.56	87.30	86.43	93.13	/
MV-Swin-T	75.63	80.62	76.82	70.63	73.83	/
CTT-Net	72.50	78.39	71.38	73.54	71.00	/
LVSnake	83.09	\	87.17	86.49	90.11	/
E-ViM³
E-ViM³ + plugin
XFMamba
XFMamba + plugin
BI-Mamba	77.50	73.07	83.85	78.12	91.59	/
BI-Mamba + plugin
DK-Mamba	86.25	91.08	88.99	90.93	87.78	85.58

6. 单切面对比实验

6.1 camus

汇总

view	name	acc	auc	f1	precision	recall
A4C	logvmamba	0.7625	0.7069	0.5456	0.8605	0.4575
	lkm_unet	0.7875	0.7537	0.5917	0.8466	0.5424
	ukan3d	0.7125	0.7374	0.6782	0.5837	0.8257
	dk-mamba	0.8063	0.7838	0.6882	0.7962	0.6181
A2C	logvmamba	0.7375	0.6521	0.5873	0.7558	0.5161
	lkm_unet	0.7250	0.7202	0.6595	0.6790	0.7454
	ukan3d	0.7063	0.7040	0.6351	0.5832	0.7060
	dk-mamba	0.7750	0.7790	0.6834	0.6880	0.6848

A4C切面

name	acc	auc	f1	precision	recall	specificity
logvmamba	0.7625	0.7069	0.5456	0.8605	0.4575	0.9299
lkm_unet	0.7875	0.7537	0.5917	0.8466	0.5424	0.9184
ukan3d	0.7125	0.7374	0.6782	0.5837	0.8257	0.6495
dk-mamba	0.8063	0.7838	0.6882	0.7962	0.6181	0.9118

dk-mamba

name	acc	auc	f1	precision	recall	specificity
Experiment_138 fold_1	0.78125	0.6999	0.6666	0.7777	0.5833	0.8999
Experiment_138 fold_2	0.78125	0.75	0.6315	0.8571	0.5	0.9499
Experiment_138 fold_3	0.90625	0.925	0.8799	0.8461	0.9166	0.8999
Experiment_138 fold_4	0.78125	0.8311	0.6315	0.75	0.5454	0.9047
Experiment_138 fold_5	0.78125	0.7142	0.6315	0.75	0.5454	0.9047
Experiment_138	0.8063	0.7838	0.6882	0.7962	0.6181	0.9118

logvmamba

name	acc	auc	f1	precision	recall	specificity
Experiment_133 fold_1	0.75	0.7250	0.6363	0.6999	0.5833	0.85
Experiment_133 fold_2	0.75	0.5833	0.5555	0.8333	0.4166	0.9499
Experiment_133 fold_3	0.84375	0.8542	0.8	0.7692	0.8333	0.85
Experiment_133 fold_4	0.75	0.6883	0.4285	1.0	0.2727	1.0
Experiment_133 fold_5	0.71875	0.6839	0.3076	1.0	0.1818	1.0
Experiment_133	0.7625	0.7069	0.5456	0.8605	0.4575	0.9299

lkm_unet

name	acc	auc	f1	precision	recall	specificity
Experiment_135 fold_1	0.875	0.8833	0.8333	0.8333	0.8333	0.8999
Experiment_135 fold_2	0.75	0.7624	0.5555	0.8333	0.4166	0.9499
Experiment_135 fold_3	0.84375	0.85	0.8148	0.7333	0.9166	0.8
Experiment_135 fold_4	0.78125	0.7273	0.5882	0.8333	0.4545	0.9523
Experiment_135 fold_5	0.6875	0.5454	0.1666	1.0	0.0909	1.0
Experiment_135	0.7875	0.7537	0.5917	0.8466	0.5424	0.9184

ukan3d

name	acc	auc	f1	precision	recall	specificity
Experiment_136 fold_1	0.75	0.8167	0.7333	0.6111	0.9167	0.6499
Experiment_136 fold_2	0.7187	0.6833	0.64	0.6153	0.6666	0.75
Experiment_136 fold_3	0.875	0.8708	0.8571	0.75	1.0	0.8
Experiment_136 fold_4	0.6875	0.7532	0.6153	0.5333	0.7272	0.6666
Experiment_136 fold_5	0.53125	0.5628	0.5454	0.4090	0.8182	0.3809
Experiment_136	0.7125	0.7374	0.6782	0.5837	0.8257	0.6495

segformer

name	acc	auc	f1	precision	recall	specificity
Experiment_137 fold_1
Experiment_137 fold_2	0.6875	0.7291	0.6666	0.5555	0.8333	0.6
Experiment_137 fold_3
Experiment_137 fold_4
Experiment_137 fold_5
Experiment_137

A2C切面

name	acc	auc	f1	precision	recall	specificity
logvmamba	0.7375	0.6521	0.5873	0.7558	0.5161	0.8629
lkm_unet	0.7250	0.7202	0.6595	0.6790	0.7454	0.7195
ukan3d	0.7063	0.7040	0.6351	0.5832	0.7060	0.7071
dk-mamba	0.7750	0.7790	0.6834	0.6880	0.6848	0.8233

dk-mamba

name	acc	auc	f1	precision	recall	specificity
Experiment_141 fold_1	0.875	0.85	0.8462	0.7857	0.9166	0.85
Experiment_141 fold_2	0.8125	0.8042	0.75	0.75	0.75	0.85
Experiment_141 fold_3	0.71875	0.7125	0.64	0.6153	0.6666	0.75
Experiment_141 fold_4	0.875	0.8268	0.8	0.8888	0.7273	0.9523
Experiment_141 fold_5	0.59375	0.7013	0.3809	0.4	0.3636	0.7142
Experiment_141	0.7750	0.7790	0.6834	0.6880	0.6848	0.8233

logvmamba

name	acc	auc	f1	precision	recall	specificity
Experiment_134 fold_1	0.6875	0.7292	0.5	0.625	0.4166	0.85
Experiment_134 fold_2	0.78125	0.6083	0.5882	1.0	0.4166	1.0
Experiment_134 fold_3	0.71875	0.6375	0.64	0.6154	0.6666	0.75
Experiment_134 fold_4	0.8125	0.5974	0.625	1.0	0.4545	1.0
Experiment_134 fold_5	0.6875	0.6883	0.5833	0.5385	0.6363	0.7143
Experiment_134	0.7375	0.6521	0.5873	0.7558	0.5161	0.8629

lkm_unet

name	acc	auc	f1	precision	recall	specificity
Experiment_139 fold_1	0.8125	0.7791	0.7272	0.8	0.6666	0.8999
Experiment_139 fold_2	0.75	0.7166	0.7333	0.6111	0.9166	0.6499
Experiment_139 fold_3	0.78125	0.7458	0.5882	1.0	0.4166	1.0
Experiment_139 fold_4	0.59375	0.7186	0.6060	0.4545	0.9090	0.4285
Experiment_139 fold_5	0.6875	0.6407	0.6428	0.5294	0.8181	0.6190
Experiment_139	0.7250	0.7202	0.6595	0.6790	0.7454	0.7195

ukan3d

name	acc	auc	f1	precision	recall	specificity
Experiment_140 fold_1	0.71875	0.6833	0.64	0.6154	0.6666	0.75
Experiment_140 fold_2	0.75	0.7583	0.7142	0.625	0.8333	0.6999
Experiment_140 fold_3	0.75	0.7625	0.6666	0.6666	0.6666	0.8
Experiment_140 fold_4	0.625	0.6883	0.5714	0.4705	0.7272	0.5714
Experiment_140 fold_5	0.6875	0.6277	0.5833	0.5384	0.6363	0.7143
Experiment_140	0.7063	0.7040	0.6351	0.5832	0.7060	0.7071

6.2 hmc

汇总

view	name	acc	auc	f1	precision	recall
A4C	logvmamba	0.7438	0.6974	0.8070	0.7626	0.8652
	lkm_unet	0.7875	0.7640	0.8520	0.7642	0.9669
	ukan3d	0.7938	0.7951	0.8569	0.7723	0.9692
	dkmamba	0.8313	0.8728	0.8625	0.8899	0.8395
A2C	logvmamba	0.75	0.69148	0.81084	0.77048	0.86518
	lkm_unet	0.75	0.7635	0.8220	0.7507	0.9114
	ukan3d	0.7563	0.7538	0.8346	0.75222	0.9540
	dkmamba	0.825	0.8574	0.8534	0.8534	0.8039

A4C切面

name	acc	auc	f1	precision	recall	specificity
logvmamba	0.7438	0.6974	0.8070	0.7626	0.8652	0.5175
lkm_unet	0.7875	0.7640	0.8520	0.7642	0.9669	0.4589
ukan3d	0.7938	0.7951	0.8569	0.7723	0.9692	0.4771
dkmamba	0.8313	0.8728	0.8625	0.8899	0.8395	0.7967

dkmamba

name	acc	auc	f1	precision	recall	specificity
Experiment_147 fold_1	0.90625	0.9468	0.9362	0.9167	0.9565	0.7777
Experiment_147 fold_2	0.84375	0.9134	0.8780	0.8999	0.8571	0.8181
Experiment_147 fold_3	0.84375	0.9020	0.8387	0.9285	0.7647	0.9333
Experiment_147 fold_4	0.75	0.7705	0.8095	0.8095	0.8095	0.6363
Experiment_147 fold_5	0.8125	0.8311	0.85	0.8947	0.8095	0.8181
Experiment_147	0.8313	0.8728	0.8625	0.8899	0.8395	0.7967

logvmamba

name	acc	auc	f1	precision	recall	specificity
Experiment_148 fold_1	0.8125	0.9082	0.8695	0.8695	0.8695	0.6666
Experiment_148 fold_2	0.875	0.8311	0.9047	0.9047	0.9047	0.8181
Experiment_148 fold_3	0.5625	0.4666	0.6111	0.5789	0.6470	0.4666
Experiment_148 fold_4	0.75	0.6320	0.8260	0.7599	0.9047	0.4545
Experiment_148 fold_5	0.71875	0.6493	0.8235	0.6999	1.0	0.1818
Experiment_148	0.7438	0.6974	0.8070	0.7626	0.8652	0.5175

lkm_unet

name	acc	auc	f1	precision	recall	specificity
Experiment_149 fold_1	0.875	0.8309	0.92	0.8518	1.0	0.5555
Experiment_149 fold_2	0.875	0.8268	0.9130	0.8399	1.0	0.6363
Experiment_149 fold_3	0.6875	0.7294	0.75	0.6521	0.8823	0.4666
Experiment_149 fold_4	0.6875	0.6017	0.8077	0.6774	1.0	0.0909
Experiment_149 fold_5	0.8125	0.8311	0.8695	0.8	0.9523	0.5454
Experiment_149	0.7875	0.7640	0.8520	0.7642	0.9669	0.4589

ukan3d

name	acc	auc	f1	precision	recall	specificity
Experiment_146 fold_1	0.875	0.9565	0.92	0.8518	1.0	0.5555
Experiment_146 fold_2	0.875	0.8398	0.9130	0.8399	1.0	0.6363
Experiment_146 fold_3	0.71875	0.7333	0.7804	0.6666	0.9411	0.4666
Experiment_146 fold_4	0.6875	0.5974	0.8077	0.6774	1.0	0.0909
Experiment_146 fold_5	0.8125	0.8484	0.8636	0.8260	0.9047	0.6363
Experiment_146	0.7938	0.7951	0.8569	0.7723	0.9692	0.4771

A2C切面

name	acc	auc	f1	precision	recall	specificity
logvmamba	0.75	0.69148	0.81084	0.77048	0.86518	0.53974
lkm_unet	0.75	0.7635	0.8220	0.7507	0.9114	0.4598
ukan3d	0.7563	0.7538	0.8346	0.75222	0.9540	0.4311
dkmamba	0.825	0.8574	0.8534	0.8534	0.8039	0.8553

dkmamba

name	acc	auc	f1	precision	recall	specificity
Experiment_145 fold_1	0.84375	0.8792	0.8837	0.9499	0.8261	0.8888
Experiment_145 fold_2	0.78125	0.8009	0.8108	0.9375	0.7143	0.9090
Experiment_145 fold_3	0.84375	0.9058	0.8387	0.9285	0.7647	0.9333
Experiment_145 fold_4	0.78125	0.8052	0.8292	0.85	0.8095	0.7272
Experiment_145 fold_5	0.875	0.8961	0.9047	0.9047	0.9047	0.8181
Experiment_145	0.825	0.8574	0.8534	0.8534	0.8039	0.8553

logvmamba

name	acc	auc	f1	precision	recall	specificity
Experiment_142 fold_1	0.84375	0.9082	0.8888	0.9090	0.8695	0.7777
Experiment_142 fold_2	0.875	0.8311	0.9047	0.9047	0.9047	0.8181
Experiment_142 fold_3	0.5625	0.4627	0.6111	0.5789	0.6470	0.4666
Experiment_142 fold_4	0.75	0.6277	0.8261	0.7599	0.9047	0.4545
Experiment_142 fold_5	0.71875	0.6277	0.8235	0.6999	1.0	0.1818
Experiment_142	0.75	0.69148	0.81084	0.77048	0.86518	0.53974

lkm_unet

name	acc	auc	f1	precision	recall	specificity
Experiment_143 fold_1	0.78125	0.7874	0.8571	0.8076	0.9130	0.4444
Experiment_143 fold_2	0.78125	0.7748	0.8444	0.7916	0.9047	0.5454
Experiment_143 fold_3	0.65625	0.6235	0.7317	0.625	0.8823	0.4
Experiment_143 fold_4	0.78125	0.8181	0.8510	0.7692	0.9523	0.4545
Experiment_143 fold_5	0.75	0.8138	0.8260	0.7599	0.9047	0.4545
Experiment_143	0.75	0.7635	0.8220	0.7507	0.9114	0.4598

ukan3d

name	acc	auc	f1	precision	recall	specificity
Experiment_144 fold_1	0.90625	0.8695	0.9333	0.9545	0.9130	0.8888
Experiment_144 fold_2	0.78125	0.7922	0.8510	0.7692	0.9523	0.4545
Experiment_144 fold_3	0.65625	0.7176	0.7555	0.6071	1.0	0.2666
Experiment_144 fold_4	0.6875	0.5670	0.7999	0.6896	0.9523	0.1818
Experiment_144 fold_5	0.75	0.8225	0.8333	0.7407	0.9523	0.3636
Experiment_144	0.7563	0.7538	0.8346	0.75222	0.9540	0.4311

7. 消融实验

7.1 camus

语义校准 二腔切面补充

name	acc	auc	f1	precision	recall	specificity
-语义校准 -二腔切面	80.00	76.98	66.13	83.45	58.87	91.14
+语义校准 -二腔切面	83.13	85.91	75.99	77.93	75.90	87.23
-语义校准 +二腔切面	81.88	83.04	73.23	76.83	70.6	88.19
+语义校准 +二腔切面	85.00	83.15	77.78	84.00	72.41	92.16

-语义校准 -二腔切面

name	acc	auc	f1	precision	recall	specificity
Experiment_150 fold_1	0.875	0.8458	0.8333	0.8333	0.8333	0.8999
Experiment_150 fold_2	0.71875	0.7666	0.4	1.0	0.25	1.0
Experiment_150 fold_3	0.8125	0.8083	0.7692	0.7142	0.8333	0.8
Experiment_150 fold_4	0.875	0.8095	0.7777	1.0	0.6363	1.0
Experiment_150 fold_5	0.71875	0.6190	0.5263	0.625	0.4545	0.8571
Experiment_150	0.8	0.7698	0.6613	0.8345	0.5887	0.9114

+语义校准 -二腔切面

name	acc	auc	f1	precision	recall	specificity
Experiment_151 fold_1	0.90625	0.9666	0.8695	0.9090	0.8333	0.9499
Experiment_151 fold_2	0.75	0.7375	0.6	0.75	0.5	0.8999
Experiment_151 fold_3	0.8125	0.8208	0.7857	0.6875	0.9166	0.75
Experiment_151 fold_4	0.84375	0.8398	0.7826	0.75	0.8181	0.8571
Experiment_151 fold_5	0.84375	0.9307	0.7619	0.8	0.7272	0.9047
Experiment_151	0.83125	0.85908	0.75994	0.77930	0.75904	0.87232

-语义校准 +二腔切面

name	acc	auc	f1	precision	recall	specificity
Experiment_152 fold_1	0.90625	0.9	0.8799	0.8461	0.9166	0.8999
Experiment_152 fold_2	0.75	0.7625	0.6363	0.6999	0.5833	0.85
Experiment_152 fold_3	0.78125	0.7666	0.6956	0.7272	0.6666	0.85
Experiment_152 fold_4	0.875	0.8658	0.8181	0.8181	0.8181	0.9047
Experiment_152 fold_5	0.78125	0.8571	0.6315	0.75	0.5454	0.9047
Experiment_152	0.81875	0.8304	0.73228	0.76826	0.706	0.88186

7.2 hmc

name	acc	auc	f1	precision	recall	specificity
-语义校准 -二腔切面	81.25	80.17	85.32	84.98	86.83	73.41
+语义校准 -二腔切面	85.62	87.37	88.52	90.44	87.20	81.98
-语义校准 +二腔切面	82.50	86.54	84.51	94.60	77.08	90.51
+语义校准 +二腔切面	86.25	91.08	88.99	90.93	87.78	85.58

-语义校准 -二腔切面

name	acc	auc	f1	precision	recall	specificity
Experiment_153 fold_1	0.90625	0.9468	0.9333	0.9545	0.9130	0.8888
Experiment_153 fold_2	0.84375	0.7878	0.8837	0.8636	0.9047	0.7272
Experiment_153 fold_3	0.8125	0.8196	0.8499	0.7391	1.0	0.60
Experiment_153 fold_4	0.75	0.7229	0.8095	0.8095	0.8095	0.6363
Experiment_153 fold_5	0.75	0.7316	0.7895	0.8824	0.7142	0.8181
Experiment_153	0.8125	0.80174	0.85318	0.84982	0.86828	0.73408

+语义校准 -二腔切面

name	acc	auc	f1	precision	recall	specificity
Experiment_154 fold_1	0.84375	0.8502	0.8888	0.9090	0.8695	0.7777
Experiment_154 fold_2	0.84375	0.9004	0.8888	0.8333	0.9523	0.6363
Experiment_154 fold_3	0.84375	0.8470	0.8484	0.875	0.8235	0.8666
Experiment_154 fold_4	0.875	0.9177	0.8947	1.0	0.8095	1.0
Experiment_154 fold_5	0.875	0.8528	0.9047	0.9047	0.9047	0.8181
Experiment_154	0.8562	0.8737	0.8852	0.9044	0.8720	0.8198

name	acc	auc	f1	precision	recall	specificity
Experiment_156 fold_1
Experiment_156 fold_2
Experiment_156 fold_3
Experiment_156 fold_4
Experiment_156 fold_5	0.71875	0.8268	0.7567	0.875	0.6666	0.8181

-语义校准 +二腔切面

name	acc	auc	f1	precision	recall	specificity
Experiment_155 fold_1	0.84375	0.8985	0.8837	0.9499	0.8260	0.8888
Experiment_155 fold_2	0.71875	0.7792	0.7567	0.875	0.6666	0.8181
Experiment_155 fold_3	0.8125	0.8745	0.7857	1.0	0.6470	1.0
Experiment_155 fold_4	0.875	0.9177	0.8947	1.0	0.8095	1.0
Experiment_155 fold_5	0.875	0.8571	0.9047	0.9047	0.9047	0.8181
Experiment_155	0.8250	0.8654	0.8451	0.9460	0.7708	0.9051

8. 序列长度实验

Backbone	Plugin	8 Frames AUC	16 Frames AUC	32 Frames AUC	Avg. Gain
E-ViM³
E-ViM³ + plugin
XFMamba
XFMamba + plugin
BI-Mamba
BI-Mamba + plugin
DK-Mamba

9. 审稿意见

Reviewer #1

Questions

• \3. Please categorize the relevance of the paper (you may choose more than one). Note the different assessment criteria for the different paper categories as outlined in the Reviewer Guidelines: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
  • MIC
• \4. How would you describe the paper?
  • Both
• \5. Please describe the main contribution of the paper.
  • This study proposes DK-Mamba, a domain knowledge-fused multi-view state space model for echocardiographic analysis of myocardial infarction. To address the bottleneck that existing Mamba models are prone to selective memory distortion and miss critical pathological features in low-signal-to-noise-ratio multi-view ultrasound videos, built upon Mamba as the backbone, the model adopts a double refinement strategy: first, the domain knowledge-guided feature refinement module injects clinical prior knowledge to calibrate the selective features; then, the stable anatomical consensus provided by the A2C view is used to perform secondary calibration on the dynamic features of the A4C view. It finally achieves state-of-the-art performance in the myocardial infarction classification task on both the CAMUS and HMC-QU datasets.
• \6. Please list the major strengths of the paper. For example, you could highlight a novel formulation, an original way to use data, a demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
  • 1.DK-Mamba deeply integrates clinical domain knowledge with multi-view anatomical consensus, and its structural design is reasonable. It solves the bottleneck problem that Mamba is prone to selective memory distortion and misses critical pathological features in low signal-to-noise ratio ultrasound videos.
    2.Experimental results show that DK-Mamba achieves state-of-the-art performance on both the CAMUS and HMC-QU myocardial infarction datasets. In particular, it achieves significant improvements in various metrics that are crucial for clinical diagnosis, and effectively reduces the missed detection rate.
• \7. Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
  • 1.The abstract contains excessive background information. Quantitative experimental results and dataset sample sizes should be added (e.g., F1-score, ACC metrics, and the quantitative improvements over existing state-of-the-art methods).
    2.Definitions of A4C and A2C should be provided, along with detailed explanations of why A4C is prone to missing tiny lesions and why A2C can provide anatomical consensus. In addition, the dataset introduction should be revised to explicitly state whether both datasets include the two views for each case.
    3.The main contributions in the introduction are ambiguous and may lead to the misunderstanding that this paper proposes fundamental improvements to the Mamba model itself. It is recommended to completely rewrite the main contributions section to clarify the actual innovations of this work.
    4.The parameter freezing status of the two pre-trained encoders (EchoPrimeVideoEncoder and EchoPrimeTextEncoder) should be clearly marked in the corresponding figures.
    5.The full name of GSC is not provided when it first appears in Figure 1, even though it is a common consensus in the field.
    6.It should be explicitly clarified whether the video feature fv by EchoPrimeVideoEncoder corresponds to the A2C view, the A4C view, or both.
    7.The authoritative source of the text prompts used for the pre-trained text encoder should be clearly stated.
    8.When calculating the cosine similarity, the rationale for using dual-view features should be explained. In addition, there is an inconsistency in the figures: Figure 1(a) does not show A2C as an input to the DKFR module, but Figure 1(b) does.
    9.Equations (3) and (4) require further explanation of their physical meanings and design motivations, rather than just presenting mathematical expressions.
    10.The disclosure of experimental settings is insufficient. The complete data preprocessing pipeline should be added, and the reason for uniformly sampling 16 frames should be explained to improve the reproducibility of the model. In addition, the authors did not mention the basic settings of the loss function or provide a brief introduction to each evaluation metric.
    11.The comparison methods in Table 1 are few and outdated. More representative state-of-the-art models from the past three years (2023-2026) should be added for fair comparison.
    12.Table 2 shows that DK-Mamba has a significantly larger parameter count and higher computational cost compared to single-view baseline models. The statement “maintaining a relatively balanced computational efficiency” in the original text needs further clarification. In addition, the computational efficiency analysis of DK-Mamba under the dual-view setting is missing.
• \8. Please rate the clarity and organization of the paper.
  • Satisfactory
• \9. Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
  • The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
• \10. Code of Ethics Check. Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation? Reviewers are expected to familiarize themselves with the MICCAI Scientific Code of Ethics before completing this assessment: https://miccai.org/index.php/about-miccai/policies/scientific-code-of-ethics/
  • No
• \13. Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
  • \4. Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
• \14. Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
  • This paper proposes DK-Mamba for myocardial infarction echocardiographic analysis by fusing domain knowledge and multi-view anatomical consensus, which alleviates Mamba’s memory distortion issue and achieves state-of-the-art performance on two datasets. These merits positively influenced my score.
    Ambiguous contributions and missing implementation details; and unsupported computational efficiency claims. These factors collectively determined my overall score.
• \16. In view of your answers above and your overall experience, how would you rate your confidence in your review?
  • Confident but not absolutely certain (3)

Reviewer #2

Questions

• \3. Please categorize the relevance of the paper (you may choose more than one). Note the different assessment criteria for the different paper categories as outlined in the Reviewer Guidelines: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

  • MIC

• \4. How would you describe the paper?

  • Application study

• \5. Please describe the main contribution of the paper.

  • This paper proposes DK-Mamba, a multi-view echocardiography model for myocardial infarction analysis. The main idea is to refine Mamba features in two stages: first through domain knowledge guidance using prompt-based semantic priors, and then through cross-view anatomical refinement using complementary echo views. The paper tries to make Mamba more reliable in noisy ultrasound videos by correcting its features with both clinical knowledge and multi-view consistency.

• \6. Please list the major strengths of the paper. For example, you could highlight a novel formulation, an original way to use data, a demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

  • The overall intuition is easy to appreciate. In echocardiography, different views often provide complementary information, so using A4C and A2C together for MI-related analysis makes a lot of sense. The idea of first using semantic priors to guide the model and then using another view to stabilize the decision is also quite natural from a clinical perspective.

    The paper has a clear story. The proposed Double Refinement framework is easy to follow conceptually: one stage tries to correct the features with domain knowledge, and the second stage tries to make them more anatomically consistent across views. This gives the method a coherent narrative, which helps the reader understand what the model is trying to do and why.

    Finally, the paper includes both quantitative results and some qualitative analysis (e.g., Grad-CAM visualizations, ablation study results).

• \7. Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

  • The paper does not fully separate the effect of the proposed method from the effect of strong external priors. In particular, the DKFR module relies on frozen EchoPrime video and text encoders, which are already powerful pretrained models. It is hard to tell how much of the improvement comes from the proposed refinement mechanism itself.

    I was also not fully convinced by the baseline selection. Several of the compared methods are not specifically designed for myocardial infarction detection in echocardiography, so the state-of-the-art claim feels a bit less strong than it first appears. I would have found the empirical case more convincing if the paper compared more directly against task-matched and domain-matched baselines.

    The use of CAMUS also raises some questions. The paper explains that CAMUS does not originally provide MI labels, so the authors added ischemia/wall-motion annotations and filtered the dataset. That is understandable from a practical point of view, but it also makes the benchmark less standard and somewhat harder to interpret in terms of reproducibility.

• \8. Please rate the clarity and organization of the paper.

  • Satisfactory

• \9. Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

  • The authors claimed to release the source code and/or dataset upon acceptance of the submission.

• \10. Code of Ethics Check. Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation? Reviewers are expected to familiarize themselves with the MICCAI Scientific Code of Ethics before completing this assessment: https://miccai.org/index.php/about-miccai/policies/scientific-code-of-ethics/

  • No

• \13. Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

  • \3. Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

• \14. Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

  • My score here was a little bit cautious. The clinical intuition is good, and I can see why the idea of combining domain knowledge with multi-view refinement is appealing. But I was not fully convinced by the experimental support: the baseline choice feels a bit uneven, the CAMUS setup is not fully standard, and it is hard to separate the effect of the proposed method from the benefit of strong external priors. I thought the idea was quite interesting, but the evidence felt less strong than the paper’s narrative.

• \16. In view of your answers above and your overall experience, how would you rate your confidence in your review?

  • Confident but not absolutely certain (3)

Reviewer #3

Questions

• \3. Please categorize the relevance of the paper (you may choose more than one). Note the different assessment criteria for the different paper categories as outlined in the Reviewer Guidelines: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
  • MIC
• \4. How would you describe the paper?
  • Methodological contribution
• \5. Please describe the main contribution of the paper.
  • This paper proposes a Domain Knowledge-Refined State Space Model for myocardial infarction (MI) diagnosis with three main contributions. First, expert-level clinical prior knowledge is incorporated into the Mamba architecture to improve pathological feature extraction robustness in low-SNR ultrasound images. Second, a Double Refinement Mechanism is introduced: semantic knowledge vectors first correct bias in Mamba’s selective states for precise MI localization, followed by cross-view anatomical consensus from the A2C pathway to further calibrate A4C dynamic features. This dual-path strategy effectively prevents key pathological features from being overlooked.
• \6. Please list the major strengths of the paper. For example, you could highlight a novel formulation, an original way to use data, a demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
  • The major strengths of this paper are as follows:
    1.The integration of expert-level clinical prior knowledge into the Mamba architecture is well-motivated and directly addresses a practical limitation of existing vision-only models in medical ultrasound analysis.
    2.The two-stage Double Refinement Mechanism effectively mitigate the risk of missing subtle pathological features under low-SNR conditions by combining semantic knowledge-guided refinement with cross-view anatomical information.
• \7. Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
  • The major weaknesses of this paper are as follows:
    \1. The overall methodological novelty of the paper is limited. The backbone relies on an existing U-shaped Mamba architecture, and the proposed additions including A2C/A4C interaction weights and textual feature fusion involve relatively straightforward operations such as simple feature concatenation, which lack sufficient technical novelty.
    \2. As observed in Fig. 1 and the dataset description, the text data contains rich semantic information beyond image inputs. However, all compared multi-view baselines only accept images as input, creating an unfair comparison. Since the proposed method benefits from additional textual modality, it should be compared against multimodal methods that similarly take both image and text as inputs.
    \3. Several compared methods, such as LoG-VMamba and LKM-UNet, are primarily designed for segmentation tasks and only accept image inputs, making them inappropriate baselines for this classification task. Given the growing body of Mamba-based multimodal classification research in recent years, the paper should include more relevant and task-aligned comparisons to better situate its contributions.
• \8. Please rate the clarity and organization of the paper.
  • Satisfactory
• \9. Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
  • The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
• \10. Code of Ethics Check. Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation? Reviewers are expected to familiarize themselves with the MICCAI Scientific Code of Ethics before completing this assessment: https://miccai.org/index.php/about-miccai/policies/scientific-code-of-ethics/
  • No
• \13. Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
  • \3. Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
• \14. Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
  • The overall score primarily reflects two major concerns. First, the methodological novelty of this paper is insufficient. The proposed framework essentially combines an existing Mamba encoder with textual feature integration during decoding, which does not constitute a substantial technical contribution. Second, the experimental comparisons are incomplete and potentially unfair, as the proposed method leverages additional textual modality while all baselines are image-only, and several compared methods are designed for segmentation rather than classification. These concerns are detailed in the major weaknesses above.
• \16. In view of your answers above and your overall experience, how would you rate your confidence in your review?
  • Very confident (4)

中文的

审稿人 #1

问题

• 3. 请对论文的相关性进行分类（可多选）。请注意不同论文类别的评估标准，详见审稿指南：https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
  • MIC
• 4. 您如何描述这篇论文？
  • 两者兼具
• 5. 请描述论文的主要贡献。
  • 本研究提出了 DK-Mamba，一种融合领域知识的多视图状态空间模型，用于心肌梗死的超声心动图分析。为解决现有 Mamba 模型在低信噪比多视图超声视频中易出现选择性记忆失真并遗漏关键病理特征的瓶颈，该模型以 Mamba 为骨干，采用双重精炼策略：首先，领域知识引导的特征精炼模块注入临床先验知识以校准选择性特征；然后，利用 A2C 切面提供的稳定解剖学共识，对 A4C 切面的动态特征进行二次校准。最终，在 CAMUS 和 HMC-QU 两个数据集的心肌梗死分类任务上均取得了最先进的性能。
• 6. 请列出论文的主要优点。例如，您可以强调新颖的公式化表述、独特的数据使用方式、临床可行性的展示、新颖的应用、特别有力的评估，或任何其他体现本工作优势的方面。请提供细节，例如，如果方法新颖，请解释哪方面新颖及其为何有趣。
  • 1. DK-Mamba 深度融合了临床领域知识与多视图解剖学共识，其结构设计合理。解决了 Mamba 在低信噪比超声视频中易出现选择性记忆失真并遗漏关键病理特征的瓶颈问题。
  • 2. 实验结果表明，DK-Mamba 在 CAMUS 和 HMC-QU 心肌梗死数据集上均达到了最先进的性能。尤其是在对临床诊断至关重要的多项指标上取得了显著提升，并有效降低了漏检率。
• 7. 请列出论文的主要缺点。请提供细节：例如，如果您指出某种公式化表述、数据使用方式、临床可行性展示或应用不具备新颖性，则必须提供先前工作的具体参考文献。
  • 1. 摘要包含过多背景信息。应添加定量实验结果和数据集样本量（例如，F1-score、ACC 指标，以及相较于现有最先进方法的定量改进幅度）。
  • 2. 应给出 A4C 和 A2C 的定义，并详细解释为何 A4C 易遗漏微小病灶，以及为何 A2C 能提供解剖学共识。此外，数据集介绍部分应修改，明确说明两个数据集是否每个病例都包含这两个切面。
  • 3. 引言中的主要贡献表述模糊，可能导致读者误以为本文对 Mamba 模型本身提出了根本性改进。建议完全重写主要贡献部分，以阐明本工作的实际创新点。
  • 4. 两个预训练编码器（EchoPrimeVideoEncoder 和 EchoPrimeTextEncoder）的参数冻结状态应在相应图中明确标示。
  • 5. GSC 的全称在图 1 首次出现时未提供，尽管这在本领域是常识。
  • 6. 应明确指出 EchoPrimeVideoEncoder 提取的视频特征 fv 是对应 A2C 切面、A4C 切面，还是两者。
  • 7. 应明确说明用于预训练文本编码器的文本提示的权威来源。
  • 8. 在计算余弦相似度时，应解释使用双视图特征的理由。此外，图片存在不一致：图 1(a) 未显示 A2C 作为 DKFR 模块的输入，但图 1(b) 显示了。
  • 9. 公式 (3) 和 (4) 需进一步解释其物理含义和设计动机，而非仅仅罗列数学表达式。
  • 10. 实验设置披露不足。应补充完整的数据预处理流程，并解释统一采样 16 帧的原因，以提高模型的可复现性。此外，作者未提及损失函数的基本设置，也未对各评估指标进行简要介绍。
  • 11. 表 1 中的对比方法较少且陈旧。应增加更多近三年（2023-2026 年）具有代表性的最新模型进行公平对比。
  • 12. 表 2 显示，与单视图基线模型相比，DK-Mamba 的参数量和计算成本明显更大。原文中“保持相对平衡的计算效率”这一表述需进一步澄清。此外，缺少 DK-Mamba 在双视图设置下的计算效率分析。
• 8. 请评价论文的清晰度和组织结构。
  • 满意
• 9. 请评价论文的可复现性。请注意，提供代码和数据是加分项，但并非接收的必要条件。
  • 投稿已提供源代码、数据集或任何其他依赖项的匿名链接。
• 10. 科研道德检查。基于您的审稿以及您对 MICCAI 科研道德规范的理解，您认为此投稿是否可能涉及潜在的道德问题或违规？审稿人应在完成此评估前熟悉 MICCAI 科研道德规范：https://miccai.org/index.php/about-miccai/policies/scientific-code-of-ethics/
  • 否
• 13. 请按 1-6 分给论文打分，6 分为最高分（6-4 分：接收；3-1 分：拒稿）。请充分利用整个分数范围。分数的分布有助于决策。
  • 4. 弱接受 —— 略高于接收阈值，但不介意被拒，取决于 rebuttal 表现
• 14. 请说明您的推荐理由。导致您给出该总分的主要因素是什么？
  • 本文提出了 DK-Mamba，通过融合领域知识和多视图解剖学共识进行心肌梗死超声心动图分析，缓解了 Mamba 的记忆失真问题，并在两个数据集上取得了最先进的性能。这些优点对我的评分产生了积极影响。
  • 贡献表述模糊，实现细节缺失，且计算效率的声明缺乏依据。这些因素共同决定了我的总体评分。
• 16. 鉴于您以上的回答及您的整体经验，您如何评价您对本次审稿的信心程度？
  • 有信心但不完全确定 (3)

审稿人 #2

问题

• 3. 请对论文的相关性进行分类（可多选）。请注意不同论文类别的评估标准，详见审稿指南：https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
  • MIC
• 4. 您如何描述这篇论文？
  • 应用研究
• 5. 请描述论文的主要贡献。
  • 本文提出了 DK-Mamba，一种用于心肌梗死分析的多视图超声心动图模型。其主要思想是分两个阶段精炼 Mamba 特征：首先通过使用基于提示的语义先验进行领域知识引导，然后利用互补的超声切面进行跨视图解剖精炼。论文试图通过临床知识和多视图一致性来修正 Mamba 在噪声超声视频中的特征，使其更加可靠。
• 6. 请列出论文的主要优点。例如，您可以强调新颖的公式化表述、独特的数据使用方式、临床可行性的展示、新颖的应用、特别有力的评估，或任何其他体现本工作优势的方面。请提供细节，例如，如果方法新颖，请解释哪方面新颖及其为何有趣。
  • 整体思路易于理解。在超声心动图中，不同切面通常提供互补信息，因此结合使用 A4C 和 A2C 进行心梗相关分析非常有意义。从临床角度看，先使用语义先验引导模型，再使用另一视图稳定决策的思路也相当自然。
  • 论文故事清晰。所提出的双重精炼框架在概念上易于理解：一个阶段尝试用领域知识修正特征，第二阶段尝试使这些特征在视图间更具解剖学一致性。这赋予了该方法连贯的叙事，有助于读者理解模型试图做什么以及为何这么做。
  • 最后，论文同时包含了定量结果和一些定性分析（例如，Grad-CAM 可视化、消融研究结果）。
• 7. 请列出论文的主要缺点。请提供细节：例如，如果您指出某种公式化表述、数据使用方式、临床可行性展示或应用不具备新颖性，则必须提供先前工作的具体参考文献。
  • 论文未能充分将所提方法的效果与强大外部先验的效果分离开。特别是，DKFR 模块依赖于冻结的 EchoPrime 视频和文本编码器，这些已经是强大的预训练模型。很难判断有多少改进是来自所提出的精炼机制本身。
  • 我也未能完全信服基线的选择。几个对比方法并非专门为超声心动图中的心肌梗死检测而设计，因此最先进的声明感觉比乍看起来要弱一些。如果论文能更直接地与任务匹配、领域匹配的基线进行比较，其实证案例会更有说服力。
  • CAMUS 数据集的使用也存在一些疑问。论文解释说 CAMUS 原本不提供心梗标签，因此作者添加了缺血/室壁运动标注并过滤了数据集。从实践角度看这可以理解，但也使得该基准不那么标准，且在可复现性方面有些难以解释。
• 8. 请评价论文的清晰度和组织结构。
  • 满意
• 9. 请评价论文的可复现性。请注意，提供代码和数据是加分项，但并非接收的必要条件。
  • 作者声明将在论文被接收后公开源代码和/或数据集。
• 10. 科研道德检查。基于您的审稿以及您对 MICCAI 科研道德规范的理解，您认为此投稿是否可能涉及潜在的道德问题或违规？审稿人应在完成此评估前熟悉 MICCAI 科研道德规范：https://miccai.org/index.php/about-miccai/policies/scientific-code-of-ethics/
  • 否
• 13. 请按 1-6 分给论文打分，6 分为最高分（6-4 分：接收；3-1 分：拒稿）。请充分利用整个分数范围。分数的分布有助于决策。
  • 3. 弱拒绝 —— 略低于接收阈值，但不介意被接收，取决于 rebuttal 表现
• 14. 请说明您的推荐理由。导致您给出该总分的主要因素是什么？
  • 我这次的评分略显谨慎。临床直觉是好的，我能理解为何将领域知识与多视图精炼相结合的想法具有吸引力。但我未能完全信服实验支持：基线选择感觉有些不均衡，CAMUS 实验设置不完全标准，且难以将所提方法的效果与强大外部先验的增益分离开。我认为这个想法相当有趣，但证据感觉不如论文叙述的那样有力。
• 16. 鉴于您以上的回答及您的整体经验，您如何评价您对本次审稿的信心程度？
  • 有信心但不完全确定 (3)

审稿人 #3

问题

• 3. 请对论文的相关性进行分类（可多选）。请注意不同论文类别的评估标准，详见审稿指南：https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
  • MIC
• 4. 您如何描述这篇论文？
  • 方法论贡献
• 5. 请描述论文的主要贡献。
  • 本文提出了一种用于心肌梗死（MI）诊断的领域知识精炼状态空间模型，主要包含三个贡献。首先，将专家级临床先验知识融入 Mamba 架构，以提高低信噪比超声图像中病理特征提取的鲁棒性。其次，引入双重精炼机制：语义知识向量首先修正 Mamba 选择性状态中的偏差以实现精准的心梗定位，随后来自 A2C 通路的跨视图解剖学共识进一步校准 A4C 动态特征。这种双路径策略有效防止关键病理特征被忽略。
• 6. 请列出论文的主要优点。例如，您可以强调新颖的公式化表述、独特的数据使用方式、临床可行性的展示、新颖的应用、特别有力的评估，或任何其他体现本工作优势的方面。请提供细节，例如，如果方法新颖，请解释哪方面新颖及其为何有趣。
  • 本文的主要优点如下：
  • 1. 将专家级临床先验知识融入 Mamba 架构的动机明确，且直接针对现有纯视觉模型在医学超声分析中的实际局限性。
  • 2. 两阶段双重精炼机制通过结合语义知识引导的精炼与跨视图解剖信息，有效降低了在低信噪比条件下遗漏细微病理特征的风险。
• 7. 请列出论文的主要缺点。请提供细节：例如，如果您指出某种公式化表述、数据使用方式、临床可行性展示或应用不具备新颖性，则必须提供先前工作的具体参考文献。
  • 本文的主要缺点如下：
  • 1. 论文的整体方法论新颖性有限。其骨干网络依赖于现有的 U 型 Mamba 架构，而所提出的新增部分，包括 A2C/A4C 交互权重和文本特征融合，涉及的是相对简单的操作，如简单的特征拼接，缺乏足够的技术新颖性。
  • 2. 如图 1 和数据集描述所示，文本数据包含了超越图像输入的丰富语义信息。然而，所有对比的多视图基线仅接受图像作为输入，造成了不公平的比较。由于所提方法受益于额外的文本模态，应与同样接受图像和文本输入的多模态方法进行比较。
  • 3. 若干对比方法，如 LoG-VMamba 和 LKM-UNet，主要设计用于分割任务且仅接受图像输入，不适合作为此分类任务的基线。鉴于近年来基于 Mamba 的多模态分类研究日益增多，论文应包含更多相关且任务对齐的比较，以更好地确立其贡献的定位。
• 8. 请评价论文的清晰度和组织结构。
  • 满意
• 9. 请评价论文的可复现性。请注意，提供代码和数据是加分项，但并非接收的必要条件。
  • 投稿已提供源代码、数据集或任何其他依赖项的匿名链接。
• 10. 科研道德检查。基于您的审稿以及您对 MICCAI 科研道德规范的理解，您认为此投稿是否可能涉及潜在的道德问题或违规？审稿人应在完成此评估前熟悉 MICCAI 科研道德规范：https://miccai.org/index.php/about-miccai/policies/scientific-code-of-ethics/
  • 否
• 13. 请按 1-6 分给论文打分，6 分为最高分（6-4 分：接收；3-1 分：拒稿）。请充分利用整个分数范围。分数的分布有助于决策。
  • 3. 弱拒绝 —— 略低于接收阈值，但不介意被接收，取决于 rebuttal 表现
• 14. 请说明您的推荐理由。导致您给出该总分的主要因素是什么？
  • 总体评分主要基于两大关切。首先，本文的方法论新颖性不足。所提出的框架本质上是将现有的 Mamba 编码器与解码过程中的文本特征集成相结合，这并未构成实质性的技术贡献。其次，实验对比不完整且可能不公平，因为所提方法利用了额外的文本模态，而所有基线均为仅图像方法，且若干对比方法是为分割而非分类设计的。这些关切已在上文的主要缺点部分详述。
• 16. 鉴于您以上的回答及您的整体经验，您如何评价您对本次审稿的信心程度？
  • 非常自信 (4)