Heterogeneity Loss to Handle Intersubject and Intrasubject Variability\n in Cancer

Abstract

Developing nations lack adequate number of hospitals with modern equipment\nand skilled doctors. Hence, a significant proportion of these nations'\npopulation, particularly in rural areas, is not able to avail specialized and\ntimely healthcare facilities. In recent years, deep learning (DL) models, a\nclass of artificial intelligence (AI) methods, have shown impressive results in\nmedical domain. These AI methods can provide immense support to developing\nnations as affordable healthcare solutions. This work is focused on one such\napplication of blood cancer diagnosis. However, there are some challenges to DL\nmodels in cancer research because of the unavailability of a large data for\nadequate training and the difficulty of capturing heterogeneity in data at\ndifferent levels ranging from acquisition characteristics, session, to\nsubject-level (within subjects and across subjects). These challenges render DL\nmodels prone to overfitting and hence, models lack generalization on\nprospective subjects' data. In this work, we address these problems in the\napplication of B-cell Acute Lymphoblastic Leukemia (B-ALL) diagnosis using deep\nlearning. We propose heterogeneity loss that captures subject-level\nheterogeneity, thereby, forcing the neural network to learn subject-independent\nfeatures. We also propose an unorthodox ensemble strategy that helps us in\nproviding improved classification over models trained on 7-folds giving a\nweighted-$F_1$ score of 95.26% on unseen (test) subjects' data that are, so\nfar, the best results on the C-NMC 2019 dataset for B-ALL classification.\n