X-Ray Anomaly Detection Models

This repository presents an ensemble based transfer learning approach for accurately classifying common thoracic diseases from Chest X-Rays (CXRs)

Dataset

We use the CheXpert dataset for training and evaluation

Models

The final ensemble consists of the following models

• DenseNet-121
• DenseNet-169
• DenseNet-201
• Inception ResNet v2
• Xception

The ensemble weights are found empirically while the disease-wise optimal prediction thresholds are found by maximizing the Younden's J Statistic

Results

The ROC curves for each individual model and the final ensemble are located here
We achieve a mean area under the curve (AUC) of 0.915 on the validation set, that comes close to the SOTA of 0.94 (at the time of writing these models, i.e., May 2020)

Illustration of a Model Pipeline Used

Illustration of the Final Ensemble

Authors

Harshit Varma, Richeek Das, Ankit Kumar Misra, Prapti Kumar