医学图像作业：Python实现的眼底血管医学图像分割实验源码+报告

# 简介 ## 一、文件目录 ```bash . │ preprocessing.py # 4步图像增强 │ get_mask.py # 获取感兴趣区域ROI二值化 │ cut_patches.py # 切成很多小图 │ models.py # 存取不同网络模型，暂时用torch写了Unet和Unet++ │ loaddata.py # 为训练测试加载数据 │ train_test_evaluate.py # 训练、预测、评估 │ merge_imgs.py # 合并预测的小patchs，合并切割后的label（因为合并切割图像与原label大小并不一样 │ finally_evaluate.py # 计算整张图像Dice和accuracy两个评价指标 │ README.md │ ├─data │ ├─CHASE │ │ ├─test │ │ │ ├─1st_manual # 4对，8张 │ │ │ └─images │ │ └─train │ │ ├─1st_manual # 10对，20张 │ │ └─images │ └─DRIVE │ ├─test │ │ ├─1st_manual # 20张 │ │ └─images │ └─train │ ├─1st_manual # 20张 │ └─images │ │ ############## 以下文件和文件夹是本人手动加进去的 ############## │ ├─show_img │ └─... # 4张实验结果展示图 │ ├─experiment_log.txt # 实验运行日志 │ │ ################ 以下数据集是运行代码后期生成的 ################ │ ├─data_process │ ├─CHASE │ │ ├─test # 3个文件夹各8张 │ │ │ ├─1st_manual │ │ │ ├─mask │ │ │ └─preprocessing │ │ └─train # 3个文件夹各20张 │ │ ├─1st_manual │ │ ├─mask │ │ └─preprocessing │ └─DRIVE │ ├─test # 3个文件夹各20张 │ │ ├─1st_manual │ │ ├─mask │ │ └─preprocessing │ └─train # 3个文件夹各20张 │ ├─1st_manual │ ├─mask │ └─preprocessing │ ├─data_crop │ ├─CHASE │ │ ├─test # 8*36*2=576张，增强图+label │ │ └─train # 20*36*2=1440张 │ └─DRIVE │ ├─test # 20*9*2=360张 │ └─train # 20*9*2=360张 │ └─result ├─CHASE │ ├─merge # 8张，合并的预测的整个图 │ ├─merge_label # 8张，合并的分割的整个label图 │ └─test # 8*36=288张，预测的单个patch图 └─DRIVE ├─merge # 20张 ├─merge_label # 20张 └─test # 20*9=180张 ``` ## 二、执行顺序 - [x] preprocessing.py：生成4步（rgb2gray，nomalized，clahe，gamma）图像增强后的图像，保存至 `./data_process/{dataset}/{train_or_test}/preprocessing/{imgname}_gamma.jpg`。 - [x] get_mask.py：生成原始图像的mask，保存至 `./data_process/{dataset}/{train_or_test}/mask/{imgname}_mask.jpg`。 - [x] cut_patches.py：先将`1st_manual`文件夹从`./data/{dataset}/{train_or_test}/`拷贝到`./data_process/{dataset}/{train_or_test}/`，然后将增强图、`1st_manual`（ground_truth）、(mask暂时没切割)分割为指定大小的小patches，分别保存至 `./data_crop/{dataset}/{train_or_test}/{index}`。 - [x] train_test_evaluate.py：调用 models.py，执行网络训练、预测、评估。models.py和loaddata.py是要被调用的两个模块。 - [x] merge_imgs.py: 合并2个数据集的predict和label。 - [x] finally_evaluate.py: 计算 Dice 和Accuracy 测度。 ```bash python preprocessing.py python get_mask.py python cut_patches.py python train_test_evaluate.py train python train_test_evaluate.py test python train_test_evaluate.py train --batch_size=8 --ckpt='weights_DRIVE_5.pth' --datasetname='DRIVE' python train_test_evaluate.py test --ckpt='weights_DRIVE_5.pth' --datasetname='DRIVE' python merge_imgs.py # 注意每次修改5处，每次合并一个数据集的预测图or标签 python merge_imgs.py python merge_imgs.py python merge_imgs.py python finally_evaluate.py ``` ## 三、实验思路 - 1.图像预处理： - 1）单通道化RGB2Gray - 2）归一化 - 3）对比度限制自适应直方图均衡化 - 4）伽马校正 - 2.图像分割成小块patch - 3.torch写网络 - Unet  - Unet++  - 4.训练与测试，计算每个小patch的train_loss和dice_score - 5.合并图像 - 6.计算整体测度 ## 四、实验结果 本人运算资源时间有限，暂时没有进行深入对比试验。`torch.__version__=1.4.0`，`torchvision.__version__=0.5.0`，GPU：天池平台的的 `Tesla V100`，显存有时不到1GB，CPU：天池平台的 `2Core4G` CHASE数据集用cuda训练batchsize为2，网络采用UNet++，轮数epoch=5，测试集结果： avarage Dice: **78.03%**, avarage Accuracy: **96.91%** DRIVE数据集用cpu训练batchsize为8，网络采用UNet，轮数epoch=5，测试集结果： avarage Dice: **77.45%**, avarage Accuracy: **95.83%** 1.预处理增强步骤图 1）单通道化RGB2Gray 2）归一化 3）对比度限制自适应直方图均衡化 CLAHE 4）伽马校正  2.切分的增强图小块图，小块标签图，小块预测图patch_labelVSpatch_predict  3.CHASE_test  4.DRIVE_test  5.合并后的图计算两个评价指标：Dice 和 Accuracy ``` CHASE No.1, Dice:82.06%, Accuracy:97.71% CHASE No.2, Dice:81.09%, Accuracy:97.62% CHASE No.3, Dice:76.20%, Accuracy:96.22% CHASE No.4, Dice:78.20%, Accuracy:96.52% CHASE No.5, Dice:77.55%, Accuracy:96.84% CHASE No.6, Dice:73.54%, Accuracy:96.46% CHASE No.7, Dice:79.55%, Accuracy:96.93% CHASE No.8, Dice:76.02%, Accuracy:97.00% CHASE avarage Dice:78.03%, avarage Accuracy:96.91% DRIVE No.1, Dice:80.14%, Accuracy:95.85% DRIVE No.2, Dice:82.00%, Accuracy:96.07% DRIVE No.3, Dice:75.77%, Accuracy:94.57% DRIVE No.4, Dice:79.41%, Accuracy:95.84% DRIVE No.5, Dice:76.71%, Accuracy:95.51% DRIVE No.6, Dice:74.82%, Accuracy:95.07% DRIVE No.7, Dice:76.64%, Accuracy:95.71% DRIVE No.8, Dice:72.05%, Accuracy:95.28% DRIVE No.9, Dice:70.27%, Accuracy:95.35% DRIVE No.10, Dice:77.66%, Accuracy:95.96% DRIVE No.11, Dice:77.46%, Accuracy:95.76% DRIVE No.12, Dice:76.08%, Accuracy:95.71% DRIVE No.13, Dice:76.95%, Accuracy:95.40% DRIVE No.14, Dice:78.46%, Accuracy:96.37% DRIVE No.15, Dice:79.59%, Accuracy:96.64% DRIVE No.16, Dice:80.35%, Accuracy:96.23% DRIVE No.17, Dice:72.79%, Accuracy:95.44% DRIVE No.18, Dice:77.54%, Accuracy:96.23% DRIVE No.19, Dice:84.87%, Accuracy:97.00% DRIVE No.20, Dice:79.51%, Accuracy:96.64% DRIVE avarage Dice:77.45%, avarage Accuracy:95.83% ``` 6.实验日志见 [./experiment_log.txt](./experiment_log.txt) ## 五、参考文献 - [1] Pizer S M, Amburn E P, Austin J D, et al. Adaptive histogram equalization and its variations[J]. Computer vision, graphics, and image processing, 1987, 39(3): 355-368. - [2] Zuiderveld K . Contrast Limited Adaptive Histogram Equalization[J]. Graphics Gems, 1994:474-485. - [3] Srinivasan S, Balram N. Adaptive contrast enhancement using local region stretching[C]//Proceedings of the 9th Asian symposium on information display. 2006: 152-155. - [4] 马青柯. 眼底图像检测和分析系统的设计与开发[D]. 暨南大学. - [5] 梁礼明, 盛校棋, 郭凯,等. 基于改进的U-Net眼底视网膜血管分割[J]. 计算机应用研究, 2020, v.37;No.342(04):293-297. - [6] Jin Q , Meng Z , Pham T D , et al. DUNet: A deformable network for retinal vessel segmentation