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	test_dsc
exp_10	0.899080

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

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

医学影像

#深度学习 #语义分割 #论文

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