RongchangZhao/CheXNet-with-localization

Weakly Supervised Learning for Findings Detection in Medical Images

CheXNet-with-localization

ADLxMLDS 2017 fall final

Team:XD

黃晴 (R06922014), 王思傑 (R06922019), 曹爗文 (R06922022), 傅敏桓 (R06922030), 湯忠憲 (R06946003)

Weakly supervised localization :

In this task, we have to plot bounding boxes for each disease finding in a single chest X-ray without goundtruth (X, Y, width, height) in training set. The workflow is shown below:

Workflow :

1. Predict findings
2. Use the classifier to plot heatmap (Grad-CAM)
3. Plot the bounding box base on Grad-CAM

Package :

Pytorch==0.2.0   torchvision==0.2.0   matplotlib   scikit-image==0.13.1   opencv_python==3.4.0.12   numpy==1.13.3  matplotlib==2.1.1  scipy==1.0.0   sklearn==0.19.1  

Environment:

• OS: Linux
• Python 3.5
• GPU: 1080 Ti
• CPU: Xeon(R) E5-2667 v4
• RAM: 500 GB

Experiments process:

1. preprocessing:

python3 preprocessing.py [path of images folder] [path to data_entry] [path to bbox_list_path] [path to train_txt] [path to valid_txt] [path of preprocessed output (folder)]

2. training:

python3 train.py [path of preprocessed output (folder)]

3. local testing:

python3 denseNet_localization.py [path to test.txt] [path of images folder]

4. Output txt format:
   After running denseNet_localization.py, you would get a txt file. The format is shown below:
   

[image_path] [number_of_detection]
[disease] [x] [y] [width] [height]
[disease] [x] [y] [width] [height]
...
[image_path] [number_of_detection]
[disease] [x] [y] [width] [height]
[disease] [x] [y] [width] [height]
...

---

For DeepQ platform testing:

upload deepQ_25.zip to the platform. Then use following command:

python3 inference.py

---

Note :

In our .py script, I used the following script to assign the task running on GPU 0.

import os
os.environ['CUDA_VISIBLE_DEVICES'] = "0"

Model :

• Image is modified from Ref [2].

Result :

Prediction

Heatmap per disease Visualization of some heat maps with its ground-truth label (red) and its prediction (blue) selected from each disease class. (From top-left to bottom: Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia and Pneumothorax)

Bounding Box per patient Visualization of some images with its ground-truth label (red) and its prediction (blue) selected from each disease class.

Refers to the report for more experiment results.

Reference:

1. ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases [Arxiv]
2. LEARNING TO DIAGNOSE FROM SCRATCH BY EXPLOITING DEPENDENCIES AMONG LABELS [Arxiv]
3. CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning [Arxiv]
4. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization [Arxiv]

Contact:

Feel free to contact me (thtang@nlg.csie.ntu.edu.tw) if you have any problem.