Learning to diagnose from scratch by exploiting dependencies among\n labels

Abstract

The field of medical diagnostics contains a wealth of challenges which\nclosely resemble classical machine learning problems; practical constraints,\nhowever, complicate the translation of these endpoints naively into classical\narchitectures. Many tasks in radiology, for example, are largely problems of\nmulti-label classification wherein medical images are interpreted to indicate\nmultiple present or suspected pathologies. Clinical settings drive the\nnecessity for high accuracy simultaneously across a multitude of pathological\noutcomes and greatly limit the utility of tools which consider only a subset.\nThis issue is exacerbated by a general scarcity of training data and maximizes\nthe need to extract clinically relevant features from available samples --\nideally without the use of pre-trained models which may carry forward\nundesirable biases from tangentially related tasks. We present and evaluate a\npartial solution to these constraints in using LSTMs to leverage\ninterdependencies among target labels in predicting 14 pathologic patterns from\nchest x-rays and establish state of the art results on the largest publicly\navailable chest x-ray dataset from the NIH without pre-training. Furthermore,\nwe propose and discuss alternative evaluation metrics and their relevance in\nclinical practice.\n