CTA数据集分割实验

数据预处理

直接读取文件夹随机按照8:2进行划分

part = 'Abdominal_Infrarenal'
part = 'Abdominal_Suprarenal'
part = 'Ascending_Arch'
part = 'Descending_Thoracic'

上述四个部位都是

1	`Total: 78, Training: 62, Validation: 16`

直接使用nii.gz文件，在最开始时，内存中直接先预处理和加载

import yaml
from monai.transforms import (
    AsDiscrete,
    EnsureChannelFirstd,
    Compose,
    CropForegroundd,
    LoadImaged,
    Orientationd,
    RandFlipd,
    RandCropByPosNegLabeld,
    RandShiftIntensityd,
    ScaleIntensityRanged,
    Spacingd,
    RandRotate90d,
    Rand3DElasticd,
    ResizeWithPadOrCropd,
    AdjustContrastd
)

config_path = "config/config.yaml"
with open(config_path, 'r', encoding='utf-8') as f:
    config = yaml.safe_load(f)

# compose将transforms组合在一起
# 图像尺寸为[1, 259, 223, 74]
train_transforms = Compose(
    [
        # 加载图片的值和元数据，参数keys是data_dicts中设置的keys，表示对image还是label做变换
        LoadImaged(keys=["image", "label", "all_lab"], image_only=False),
        # 自动添加一个通道的维度，保证通道在第一维度
        EnsureChannelFirstd(keys=["image", "label", "all_lab"]),
        # 对图像进行一个方向变换，转为RAS坐标
        Orientationd(keys=["image", "label", "all_lab"], axcodes="RAS"),
        # 对图像进行重采样，体素间距重采样为[1.5, 1.5, 2.0]
        Spacingd(
            keys=["image", "label", "all_lab"],
            pixdim=config['transforms']['spacing']['pixdim'],
            mode=(config['transforms']['spacing']['mode']),
        ),
        # 对图像值强度进行归一化，由a的范围归一到b，，不在a范围的值设置为0
        ScaleIntensityRanged(
            keys=["image"],
            a_min=config['transforms']['scale_intensity']['a_min'],
            a_max=config['transforms']['scale_intensity']['a_max'],
            b_min=config['transforms']['scale_intensity']['b_min'],
            b_max=config['transforms']['scale_intensity']['b_max'],
            clip=config['transforms']['scale_intensity']['clip'],
        ),
        # 根据key所指定的大于0的部分，对于图像的有效部分进行裁剪
        CropForegroundd(keys=["image", "label", "all_lab"], source_key="label",
                        margin=config['transforms']['crop_foreground']['margin']),
        ResizeWithPadOrCropd(keys=["image", "label", "all_lab"], spatial_size=config['transforms']['resize']['spatial_size'],
                             mode=config['transforms']['resize']['mode']),
        # AdjustContrastd(keys=["image"], gamma = 2),
        # 将图像裁剪为4个子图
        RandCropByPosNegLabeld(
            keys=["image", "label", "all_lab"],
            label_key="label",
            spatial_size=config['transforms']['rand_crop']['spatial_size'],
            pos=config['transforms']['rand_crop']['pos'],
            neg=config['transforms']['rand_crop']['neg'],  # 设置为0，因为我们没有背景样本
            num_samples=config['transforms']['rand_crop']['num_samples'],
            image_key="image",
            image_threshold=0,
        ),
        # 随机旋转，按着0轴旋转，旋转概率为0.1
        RandFlipd(
            keys=["image", "label", "all_lab"],
            spatial_axis=[0],
            prob=0.5,
        ),
        # 随机旋转，按着1轴旋转，旋转概率为0.1
        RandFlipd(
            keys=["image", "label", "all_lab"],
            spatial_axis=[1],
            prob=0.5,
        ),
        # 随机旋转，按着2轴旋转，旋转概率为0.1
        RandFlipd(
            keys=["image", "label", "all_lab"],
            spatial_axis=[2],
            prob=0.5,
        ),
        # 随机旋转，概率为0.1，旋转次数为3
        RandRotate90d(
            keys=["image", "label", "all_lab"],
            prob=0.5,
            max_k=4,
        ),
        # 随机强度转换，强度偏移量为[-0.1, 0.1]，概率为0.5
        RandShiftIntensityd(
            keys=["image"],
            offsets=config['transforms']['rand_shift_intensity']['offsets'],
            prob=config['transforms']['rand_shift_intensity']['prob'],
        ),
    ]
)

val_transforms = Compose(
    [
        # 加载图片的值和元数据，参数keys是data_dicts中设置的keys，表示对image还是label做变换
        LoadImaged(keys=["image", "label", "all_lab"], image_only=False),
        # 自动添加一个通道的维度，保证通道在第一维度
        EnsureChannelFirstd(keys=["image", "label", "all_lab"]),
        # 对图像进行一个方向变换，转为RAS坐标
        Orientationd(keys=["image", "label", "all_lab"], axcodes="RAS"),
        # 对图像进行重采样，体素间距重采样为[1.5, 1.5, 2.0]
        Spacingd(
            keys=["image", "label", "all_lab"],
            pixdim=config['transforms']['spacing']['pixdim'],
            mode=(config['transforms']['spacing']['mode']),
        ),
        # 对图像值强度进行归一化，由a的范围归一到b，，不在a范围的值设置为0
        ScaleIntensityRanged(
            keys=["image"],
            a_min=config['transforms']['scale_intensity']['a_min'],
            a_max=config['transforms']['scale_intensity']['a_max'],
            b_min=config['transforms']['scale_intensity']['b_min'],
            b_max=config['transforms']['scale_intensity']['b_max'],
            clip=config['transforms']['scale_intensity']['clip'],
        ),
        # 根据key所指定的大于0的部分，对于图像的有效部分进行裁剪
        CropForegroundd(keys=["image", "label", "all_lab"], source_key="label",
                        margin=config['transforms']['crop_foreground']['margin']),
        ResizeWithPadOrCropd(keys=["image", "label", "all_lab"], spatial_size=config['transforms']['resize']['spatial_size'],
                             mode=config['transforms']['resize']['mode']),
    ]
)

增加DIS部位

DIS

1	`Total: 56, Training: 44, Validation: 12`

DIS_Abdominal_Infrarenal

1	`Total: 29, Training: 23, Validation: 6`

DIS_Abdominal_Suprarenal

1	`Total: 30, Training: 24, Validation: 6`

DIS_Ascending_Arch

1	`Total: 36, Training: 28, Validation: 8`

DIS_Descending_Thoracic

1	`Total: 46, Training: 36, Validation: 10`

实验

exp	test_dsc	test_iou
V2	0.725165	0.568831

使用basicunet

将img和all_lab在通道维度进行拼接

Abdominal_Infrarenal

exp	test_dsc
exp_2	0.758985

Abdominal_Suprarenal

exp	test_dsc
exp_3	0.823703

Ascending_Arch部位

exp	test_dsc
exp_1	0.884087

Descending_Thoracic

exp	test_dsc
exp_5	0.700073
exp_6	0.656711

DIS

exp	test_dsc
exp_17	0.818056

DIS_Abdominal_Infrarenal

exp	test_dsc
exp_18	0.701003

DIS_Abdominal_Suprarenal

exp	test_dsc
exp_19	0.836815

DIS_Ascending_Arch

exp	test_dsc
exp_20	0.821586

DIS_Descending_Thoracic

exp	test_dsc
exp_21	0.672169

使用segmamba

Abdominal_Infrarenal

exp	test_dsc
exp_8	0.768712

Abdominal_Suprarenal

exp	test_dsc
exp_9	0.847736

Ascending_Arch部位

exp	dsc	iou
exp_10	0.899080
exp_260401-1027	0.8518	0.8032

Descending_Thoracic

exp	test_dsc
exp_7	0.755863

DIS

exp	test_dsc
exp_22	0.819693

DIS_Abdominal_Infrarenal

exp	test_dsc
exp_23	0.633685
exp_31	0.893325

DIS_Abdominal_Suprarenal

exp	test_dsc
exp_24	0.884835

DIS_Ascending_Arch

exp	test_dsc
exp_25	0.872596

DIS_Descending_Thoracic

exp	test_dsc
exp_26	0.828840

使用echomamba

Abdominal_Infrarenal

exp	test_dsc
exp_11	0.763986

Abdominal_Suprarenal

exp	test_dsc
exp_12	0.797840
exp_16	0.797988

Ascending_Arch部位

exp	test_dsc
exp_14	0.878406

Descending_Thoracic

exp	test_dsc
exp_15	0.730438

DIS

exp	test_dsc
exp_26	0.754187

DIS_Abdominal_Infrarenal

exp	test_dsc
exp_27	0.843867

DIS_Abdominal_Suprarenal

exp	test_dsc
exp_28	0.926013

DIS_Ascending_Arch

exp	test_dsc
exp_29	0.799774
exp_32	0.929628

DIS_Descending_Thoracic

exp	test_dsc
exp_30	0.809732

使用nnunet

Abdominal_Infrarenal

exp	test_dsc	test_cldsc
exp_48	64.49	64.38

Abdominal_Suprarenal

exp	test_dsc	test_cldsc
exp_49	61.15	55.88
	out1

Ascending_Arch部位

exp	test_dsc	test_cldsc
exp_49	83.47	71.47

Descending_Thoracic

exp	test_dsc	test_cldsc
exp_49	71.37	69.96

DIS

exp	test_dsc	test_cldsc
exp_49	75.42	72.77

DIS_Abdominal_Infrarenal

exp	test_dsc	test_cldsc
exp_49	65.25	60.18

DIS_Abdominal_Suprarenal

exp	test_dsc	test_cldsc
exp_49	61.47	54.40

DIS_Ascending_Arch

exp	test_dsc	test_cldsc
exp_49	85.45	80.53

DIS_Descending_Thoracic

exp	test_dsc	test_cldsc
exp_50	71.74	59.20

前期汇总

	U-net	EchoMamba	SegMamba
Abdominal_Infrarenal	75.90	76.40	76.87
Abdominal_Suprarenal	82.37	79.80	84.77
Ascending_Arch	88.41	87.84	89.91
Descending_Thoracic	70.01	73.04	75.59
DIS	81.81	75.42	81.97
DIS_Abdominal_Infrarenal	70.10	84.39	89.33
DIS_Abdominal_Suprarenal	83.68	92.60	88.48
DIS_Ascending_Arch	87.26	92.96	87.26
DIS_Descending_Thoracic	67.22	80.97	82.88

nnunet	Dice	clDice
Abdominal_Infrarenal	64.49	64.38
Abdominal_Suprarenal	61.15	55.88
Ascending_Arch	83.47	71.47
Descending_Thoracic	71.37	69.96
DIS	75.42	72.77
DIS_Abdominal_Infrarenal	65.25	60.18
DIS_Abdominal_Suprarenal	61.47	54.40
DIS_Ascending_Arch	85.45	80.53
DIS_Descending_Thoracic	71.74	59.20

hilbert3d_mamba

Ascending_Arch部位

exp	dsc	iou	cldsc
exp_260331-1820	0.7321	0.6600	0.1841
exp_260331-1820	0.8059	0.7408	0.3574
exp_260331-1925	0.7039	0.6329	0.1282
exp_260331-2034	0.8439	0.7899	0.6961

hilbert3d_mamba	exp	Dice	IOU
Abdominal_Infrarenal	exp_260331-1136	0.7239	0.6782
Abdominal_Suprarenal	exp_260331-1244	0.6697	0.6124
Ascending_Arch	exp_260331-1350	0.8294	0.7695
Descending_Thoracic	exp_260331-1441	0.7356	0.6729
DIS	exp_260331-1605	0.7196	0.6586
DIS_Abdominal_Infrarenal
DIS_Abdominal_Suprarenal
DIS_Ascending_Arch
DIS_Descending_Thoracic

小方向

neck数据集

label_1 LSTA

model	dice	cldice
unet	49.99	/
segmamba	49.99	/
EM-Mamba	49.99	/
nnunet	62.98	32.67

label_2 RSTA

model	dice	cldice
unet	49.99	/
segmamba	50.18	0.05
EM-Mamba	50.23	0.05
nnunet	63.35	33.23

label_3 STA

model	dice	cldice
unet	50.19	0.05
segmamba	50.20	0.05
EM-Mamba	50.06	/
nnunet	75.73	60.75

Upper_limb_vessels

model	dice	cldice
unet	45.88	37.50
segmamba	66.09	57.37
EM-Mamba	70.42	60.26
nnunet	77.14	67.23

AI-ERCP

model	dice	iou	cldice
unet e85	0.7178	0.6364	0.7435
segformer3d e83	0.6408	0.5725	0.6638
segmamba e86	0.6927	0.6140	0.7187
EM-Mamba e84	0.7182	0.6364	0.7355
LKM-Unet e78	0.7375	0.6538	0.7792
LoG-VMamba e80	0.7258	0.6433	0.7725
LightMUnet e88	0.6599	0.5872	0.6929
UKAN3D e81	0.7311	0.6497	0.7581
nnformer e92	0.7088	0.6271	0.7246
nnunet e91	0.7400	0.6557	0.7683
unetr++ e93	0.7258	0.6435	0.7610
transunet e94	0.7318	0.6487	0.7637

nnformer

需要[128 128 128]

model	dice	iou	cldice
e92	0.7088	0.6271	0.7246

unetr++

model	dice	iou	cldice
e93	0.7258	0.6435	0.7610

nnunet

model	dice	iou	cldice
e91	0.7400	0.6557	0.7683

transunet

model	dice	iou	cldice
e94	0.7318	0.6487	0.7637

unet

model	dice	iou	cldice
e54	0.6424	0.5851	0.6488
e85	0.7178	0.6364	0.7435

segformer3d

model	dice	iou	cldice
e63	0.5293	0.5124	0.5350
e70	0.5753	0.5382	0.5845
e71	0.5641	0.5291	0.5766
e83	0.6408	0.5725	0.6638
e87	0.6327	0.5667	0.6616

segmamba

model	dice	iou	cldice
e56	0.6014	0.5572	0.6026
e72	0.6043	0.5566	0.6190
e73	0.6203	0.5686	0.6257
e86	0.6927	0.6140	0.7187

EM-Mamba

model	dice	iou	cldice
e57	0.5904	0.5510	0.5966
e74	0.6035	0.5562	0.6204
e75	0.6181	0.5667	0.6197
e84	0.7182	0.6364	0.7355

LKM-Unet

model	dice	iou	cldice
e58	0.6470	0.5903	0.6372
e77	0.7277	0.6457	0.7619
e78	0.7375	0.6538	0.7792

LoG-VMamba

model	dice	iou	cldice
e64	0.6004	0.5584	0.5859
e79	0.7235	0.6410	0.7626
e80	0.7258	0.6433	0.7725

LightMUnet

model	dice	iou	cldice
e62	/	/	/
e88	0.6599	0.5872	0.6929
e89	0.6341	0.5671	0.6747
e90	0.6438	0.5743	0.6812

UKAN3D

model	dice	iou	cldice
e61	0.6530	0.5926	0.6545
e81	0.7311	0.6497	0.7581
e82	0.7247	0.6435	0.7576

hilbert3d mamba

model	dice	iou	cldice
exp_260330-1722	0.5770	0.5415	0.1791
exp_260330-1747	0.7136	0.6367	0.4191
exp_260330-1923	0.7118	0.6348	0.4212
exp_260330-1927	0.5883	0.5468	0.2100

使用新的代码处理的数据集（不如之前那个

model	dice	iou	cldice
unet e59	0.5080	0.5031	0.5077
UKAN3D e60	0.6423	0.5855	0.6444

新的数据预处理

# 增加清晰度
pixdim: [0.8, 0.8, 1.5]
spatial_size: [160, 160, 96]
spatial_size: [64, 64, 64]
# 调整HU
a_min: -20
a_max: 250

model	dice	iou	cldice
unet e65	0.6009	0.5514	0.6261
UKAN3D e66	0.5843	0.5412	0.5966

# 增加清晰度
pixdim: [0.8, 0.8, 1.5]
spatial_size: [160, 160, 96]
spatial_size: [96, 96, 96]
# 调整HU
a_min: -20
a_max: 250

model	dice	iou	cldice
unet e67	0.5444	0.5173	0.5503

# 增加清晰度
pixdim: [0.8, 0.8, 1.5]
spatial_size: [160, 160, 96]
spatial_size: [64, 64, 64]
# 调整HU
a_min: -125
a_max: 250

model	dice	iou	cldice
unet e68	0.6037	0.5534	0.6206

加上上述调整的情况下，调整正负比例

1 2	`pos: 3 neg: 1`

model	dice	iou	cldice
unet e69	0.5209	0.5048	0.5246

resize:
#   spatial_size: [192, 128, 96]
    spatial_size: [512, 512, 96]
    mode: "constant"

model	dice	iou	cldice
unet e76	0.5552	0.5273	0.5412

CA分割

主动脉根部复合体

model	dice	iou	cldice
nnunet e95	0.7182	0.6374	0.6462
unet e99	0.8112	0.7305	0.6758
unet e105	0.7347	0.6560	0.6680
unet e106	0.6590	0.5936	0.2441

新处理的ca

model	dice	iou	cldice
unet e107	0.7192	0.6407	0.3060
unet e108	0.8343	0.7573	0.6056
unet e110	0.8377	0.7620	0.4592
unet e111	0.5924	0.5506	0.0501
unet e112	0.8158	0.7387	0.6489
unet e113	0.8918	0.8270	0.4644
unet e114	0.8782	0.8098	0.5814

CAC分割

主动脉钙化组织

model	exp	dice	iou	cldice
nnunet	e96	0.5483	0.5483	0.5483
nnunet	e97	0.5489	0.5486	0.5493
nnunet	e98	0.5502	0.5493	0.5
unet	e100	0.5484	0.5481	0.5
unet	e101	0.5650	0.5573	0.5618
unet	e102	0.5487	0.5485	0.5489
unet	e103	0.5545	0.5516	0.5533
unet	e104	0.5486	0.5485	0.5494
unet	e109	0.5969	0.5968
unet	e115	0.5967	0.5968	0.1935
unet	e116	0.5973	0.5970	0.1943
unet	e117	0.5967	0.5968	0.1935
hilbert3d_mamba	exp_260330-1654	0.5967	0.5968	0.1935

医学影像

#深度学习 #论文 #语义分割

CTA数据集分割实验

http://example.com/2025/09/11/CTA-seg/

作者

Mercury

发布于

2025年9月11日

许可协议

mamba心脏疾病诊断上一篇

RocketMQ 下一篇