CheXNet for Classification and Localization of Thoracic Diseases
This is a Python3 (Pytorch) reimplementation of CheXNet. The model takes a chest X-ray image as input and outputs the probability of each thoracic disease along with a likelihood map of pathologies.
The ChestX-ray14 dataset comprises 112,120 frontal-view chest X-ray images of 30,805 unique patients with 14 disease labels. To evaluate the model, we randomly split the dataset into training (70%), validation (10%) and test (20%) sets, following the work in paper. Partitioned image names and corresponding labels are placed under the directory labels.
- Python 3.4+
- PyTorch and its dependencies
Clone this repository.
Specify one or multiple GPUs and run
We followed the training strategy described in the official paper, and a ten crop method is adopted both in validation and test. Compared with the original CheXNet, the per-class AUROC of our reproduced model is almost the same. We have also proposed a slightly-improved model which achieves a mean AUROC of 0.847 (v.s. 0.841 of the original CheXNet).
|Pathology||Wang et al.||Yao et al.||CheXNet||Our Implemented CheXNet||Our Improved Model|
This work was collaboratively conducted by Xinyu Weng, Nan Zhuang, Jingjing Tian and Yingcheng Liu.
All of us are students/interns of Machine Intelligence Lab, Institute of Computer Science & Technology, Peking University, directed by Prof. Yadong Mu (http://www.muyadong.com).