Task Overview In this task, the objective is to identify the presence of 14 diseases in a given chest X-ray image. Chest X-ray14 dataset is the first largescale dataset on 14 common diseases in chest X-rays. The dataset contains 112,120 images from 30,805 unique patients. Image-level labels are mined from image attached reports using natural language processing techniques (each image can have multi-labels).
Download the cxr14 dataset from https://nihcc.app.box.com/v/ChestXray-NIHCC and un-tar all files. You can specify where the data is located in the run script with the below option.
--path_to_images <PATH_TO_CXR14_DATA>
You may start with the below docker file and install additional packages.
FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-develpip install -r requirements.txtOur code uses wandb for tracking metrics and losses. So you must login before starting the run.
wandb login <YOUR_WANDB_API_KEY>
To train resnet50-ACM from scratch, you may use the below code.
sh scripts/train.sh
If you just want to evaluate the trained model, you may use the below script. You may download the pretrained model from here.
sh scripts/eval.sh
| label | auc |
|---|---|
| Atelectasis | 0.8342763556741988 |
| Cardiomegaly | 0.9071881856141923 |
| Consolidation | 0.8087410332298132 |
| Edema | 0.9021789678229688 |
| Effusion | 0.8866042854223145 |
| Emphysema | 0.9482558484970659 |
| Fibrosis | 0.8508758545762228 |
| Hernia | 0.9477317680086773 |
| Infiltration | 0.7193759339295139 |
| Mass | 0.8628153089659782 |
| Nodule | 0.8153894291376429 |
| Pleural_Thickening | 0.8007042676242088 |
| Pneumonia | 0.7725149814338228 |
| Pneumothorax | 0.8976015387655567 |
This code is refactored from https://github.com/jrzech/reproduce-chexnet