Explanation Methods for a Medical Image Classifier by Analysis of its Uncertainty

Abstract: Over the last decade, neural networks have reached almost every field of science and technology. They have become a crucial part of various real-world applications, such as medical imaging. Still, their deployment in safety-critical applications remains limited owing to their inability to provide reliable uncertainty estimates and frequently occurring overconfident predictions, which is normally the case in modern neural networks possessing a substantial number of layers. In this thesis, we leverage the capability of data mining algorithms like density clustering to explain the behavior of a medical image classifier responsible for classifying white blood cells. We know that any clustering algorithm acts on the feature vector of the input data and annotates the data into different clusters as per the features. In this work, we lay down and prove the hypothesis that the output discrete probability matrix of a multi-class classification problem can be used as a feature vector where the confidence value of every class can be considered as a degree of resemblance with that class. Before implementing clustering, one needs to make sure that these confidence values represent actual probabilities so that they can be used as features; hence certain calibration techniques were incorporated to improve the calibration of the network first. Having a better calibrated medical classifier, density clustering was implemented, which generated results that provided solid arguments to justify the behavior of the network. As far as the use case of this method is concerned, it was observed that we could identify pathologies like myelodysplastic syndromes, acute lymphocytic leukemia, and chronic myelomonocytic leukemia in a patient. This was possible due to the presence of the same class of White blood cells in multiple clusters indicating the presence of subpopulations separated into healthy and pathological cells of the same class depending upon the pathology that needs to be detected. This was proved visually by mapping cluster points to actual cell images and quantitatively as well by using entropy as a method of quantifying uncertainty. This method showed that there is a lot of information embedded in the output probability matrix. Hence one can employ various data mining techniques to extract more information and not just limit themselves to misclassifications and confusion matrices.