EXPLAINABLE CLASSIFICATION OF X-RAY MEDICAL IMAGES USING BAYESIAN DEEP LEARNING

This repository contains the code base for the project for the Advanced Machine Learning (02460) course at DTU.

Please refer to our paper here for further and detailed explantions.

The collaborators were:

• Grigor Spalj
• Alessandro Contini
• Anders David Lægdsgaard Lassen
• Mads Birch Sørensen

ABSTRACT

Deep Neural Networks (DNNs) have often been found to be poorly calibrated due to overconfident predictions. At the same time the increasing complexity of modern DNNs have led to these models being considered black boxes. For that reason, various explanation methods have been proposed to uncover what features influence the predictions of DNNs. Bayesian Deep Neural Networks (BNNs) infer the entire posterior distribution of the weights, meaning that uncertainty about predictions is inherent. In this code we investigate how a Bayesian approach can improve the calibration and explainability of modern DNNs. We implemented and trained a Convolutional Neural Network (CNN) on the MURA dataset to perform a classification task and found that the Bayesian framework resulted in a significant reduction in the calibration error and improved the interpretability of the implemented visual explanation methods.

RESULTS

We found that adopting a Bayesian framework, resulted in a dramatic decrease in calibration error, indicating that the models confidence more closely matches the actual uncertainty of the predictions.

Fig. 1 - Accuracy and calibration error for the MAP and Laplace and Ensemble models.

By visually investigating the predictions of the models using Class Activation Maps (CAMs), we discovered that the models had learned optimal but different solutions. By looking at the diverse set of learned representation, valuable information about the certainty of predictions was gained.

Fig. 2 - Individual CAMs for each of the five models in the ensemble. Each column is a CAM for a single model for a hand and elbow respectively.

By aggregating the CAMs using the average,union and intersection methods, we further investigated uncertainty of the model predictions. By investigating what information is most important for the models when making predictions, we gained valuable insight into the certainty of the learned representations, and thereby about the certainty about the predictions of the models. Hence, by adopting a Bayesian framework we successfully reduced the black-box nature of modern CNNs.

Fig. 3 - Aggregated CAMs for the models in the ensemble using different aggregation methods.