AI Praxis Coursework

人工智能实践小组作业

Environment

Prerequisites

Package

pytorch
torchvision
numpy
matplotlib
pandas
seaborn
scikit-learn
opencv
# for segmentation tasks
segmentation-models-pytorch
albumentations

Create conda env

conda create --name env_name --file requirements.txt
conda activate env_name
conda install pytorch torchvision cudatoolkit=11.3 -c pytorch
pip install opencv-python
conda install ipykernel
python -m ipykernel install --user --name env_name --display-name "Python (env_name)"
conda install jupyter

Install packages for segmentation tasks

conda activate env_name
conda install segmentation-models-pytorch albumentations -c conda-forge --yes

File List

模型训练需用到的文件

+---Data
|   +---ISIC2017
|   |   +---Aug_Training_Data			     	# 扩充训练集（原始+Extra）
|   |   +---ISIC-2017_Test_Data					# 原始测试集
|   |   +---ISIC-2017_Test_v2_Part1_GroundTruth		# 测试集masks
|   |   +---ISIC-2017_Training_Part1_GroundTruth	# 训练集masks
|   |   +---ISIC-2017_Validation_Data			# 原始验证集
|   |   +---ISIC-2017_Validation_Part1_GroundTruth	# 验证集masks
|   |   +---ISIC-2017_Test_v2_Part3_GroundTruth.csv
|   |   +---ISIC-2017_Training_Aug_Part3_GroundTruth.csv
|   |   \---ISIC-2017_Validation_Part3_GroundTruth.csv
+---model
|   |   +---arl18.pth				# arl18（ARL-UNet预训练参数）
|   |   \---resnet50-19c8e357.pth		# resnet50 （ARL50预训练参数）
+---nets
|   |   +---__init__.py
|   |   +---arl.py
|   |   +---arl_unet.py
|   |   +---res_unet.py
|   |   +---resnet.py
|   |   \---unet.py
+---utils
|   |   +---__init__.py
|   |   +---data.py			# 数据加载
|   |   +---evaluation.py		# 模型评估
|   |   +---logger.py			# 日志
|   |   +---loss.py			# 损失函数
|   |   +---model.py			# 模型加载、存储、训练，checkpoint
|   |   \---visualize.py		# plot 样本、混淆矩阵
+---ARL-CNN.ipynb			# 分类模型
+---Evaluation-Cls.ipynb			# 分类模型性能评估
+---Evaluation-Seg.ipynb			# 分割模型性能评估
+---Segmentation.ipynb			# 分割模型
\---seg_main.py				# 分割模型训练

预训练模型下载

# resnet50
wget https://download.pytorch.org/models/resnet50-19c8e357.pth

Dataset

ISIC 2017: ISIC Data

分割任务mask

https://isic-challenge-data.s3.amazonaws.com/2017/ISIC-2017_Training_Part1_GroundTruth.zip
https://isic-challenge-data.s3.amazonaws.com/2017/ISIC-2017_Validation_Part1_GroundTruth.zip
https://isic-challenge-data.s3.amazonaws.com/2017/ISIC-2017_Test_v2_Part1_GroundTruth.zip

Experimental Settings and Results

Classification Task

An unofficial implementation of Attention Residual Learning for Skin Lesion Classification¹.

1. Attention Mechanism

   • 1-1 ARL-50
   • 1-2 ResNet-50

   Network Architecture	Melanoma				Seborrheic Keratosis
   	AUC	ACC	Sensitivity	Specificity	AUC	ACC	Sensitivity	Specificity
   ARL-CNN14	0.882	0.857	0.632	0.911	0.955	0.925	0.811	0.945
   ARL-CNN50	0.861	0.853	0.590	0.917	0.959	0.927	0.722	0.963
   Baseline (ARL-CNN50)	0.875	0.850	0.658	0.896	0.958	0.868	0.878	0.867
   ResNet50	0.849	0.860	0.504	0.946	0.919	0.908	0.533	0.975
   Baseline (ResNet50)	0.857	0.838	0.632	0.888	0.948	0.842	0.867	0.837

2. Loss Function

   Network: ARL-14

   • Focal Loss
     • 2-1 alpha = [1, 1, 1]
     • 2-2 alpha = [2.5, 2.5, 0.5]
   • Entropy Loss
     • 2-3 weights = [1, 1, 1]
     • 2-4 weights = [2.5, 2.5, 0.5]

   ID	Loss Function	Class Weights	Melanoma				Seborrheic Keratosis
   			AUC	ACC	Sensitivity	Specificity	AUC	ACC	Sensitivity	Specificity
   2-1	Focal Loss	[1, 1, 1]	0.846	0.858	0.538	0.936	0.924	0.912	0.511	0.982
   2-2		[2.5, 2.5, 0.5]	0.863	0.862	0.556	0.936	0.927	0.903	0.444	0.984
   2-3*	CrossEntropy Loss	[1, 1, 1]	0.822	0.835	0.487	0.919	0.929	0.893	0.833	0.904
   2-4*		[2.5, 2.5, 0.5]	0.882	0.857	0.632	0.911	0.955	0.925	0.811	0.945
   Baseline (ARL-CNN50)			0.875	0.850	0.658	0.896	0.958	0.868	0.878	0.867

`*: ran on RTX 3090, while others ran on A30.

Segmentation Task

Semantic segmentation on ISIC 2017 based on U-Net: Convolutional Networks for Biomedical Image Segmentation².

The down-sampling conv layers are changed into blocks in experiment 2~5.

ID	Settings		Results
	Network Architecture	Pretrained	ACC	SE	SP	JSI	DSC	TJI
1	UNet	F	0.8924	0.6872	0.9705	0.7525	0.8539	0.6784
2	Dense-UNet	F	0.9215	0.8224	0.9593	0.8207	0.8995	0.7420
3	Dense-UNet	T	0.9335	0.8452	0.9672	0.8457	0.9150	0.7875
4	Res-UNet	T	0.9222	0.7765	0.9778	0.8174	0.8972	0.7447
5	ARL-UNet	T	0.9220	0.7795	0.9764	0.8173	0.8972	0.7413

Tips

Run in a terminal

# (activate env)
conda activate env_name
# convert .ipynb to .py
jupyter nbconvert --to script ARL-CNN.ipynb
# nohup
nohup python ARL-CNN.py &

Run with argparse

## 分割实验2
python seg_main.py -c=0(or 1) -p=0

Check GPU info

nvidia-smi

Unzip

unzip ISIC-2017_Training_Part1_GroundTruth.zip
tar -xzvf Training_Patch.tar.gz

References

Footnotes

1. Zhang, Jianpeng, Yutong Xie, Yong Xia, and Chunhua Shen. ‘Attention Residual Learning for Skin Lesion Classification’. IEEE Transactions on Medical Imaging 38, no. 9 (September 2019): 2092–2103. https://doi.org/10.1109/TMI.2019.2893944. ↩

2. Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. ‘U-Net: Convolutional Networks for Biomedical Image Segmentation’. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, edited by Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi, 9351:234–41. Lecture Notes in Computer Science. Cham: Springer International Publishing, 2015. https://doi.org/10.1007/978-3-319-24574-4_28. ↩