2020/04/24 by Arka Mitra, Mitra, Arka, Arunava Chakravarty +9
Medicine · #COVID-19 diagnosis using AI #Lung Cancer Diagnosis and Treatment #Radiomics and Machine Learning in Medical Imaging
paper · pdf · doi:10.48550/arxiv.2004.11693
Chest radiographs are primarily employed for the screening of pulmonary and\ncardio-/thoracic conditions. Being undertaken at primary healthcare centers,\nthey require the presence of an on-premise reporting Radiologist, which is a\nchallenge in low and middle income countries. This has inspired the development\nof machine learning based automation of the screening process. While recent\nefforts demonstrate a performance benchmark using an ensemble of deep\nconvolutional neural networks (CNN), our systematic search over multiple\nstandard CNN architectures identified single candidate CNN models whose\nclassification performances were found to be at par with ensembles. Over 63\nexperiments spanning 400 hours, executed on a 11:3 FP32 TensorTFLOPS compute\nsystem, we found the Xception and ResNet-18 architectures to be consistent\nperformers in identifying co-existing disease conditions with an average AUC of\n0.87 across nine pathologies. We conclude on the reliability of the models by\nassessing their saliency maps generated using the randomized input sampling for\nexplanation (RISE) method and qualitatively validating them against manual\nannotations locally sourced from an experienced Radiologist. We also draw a\ncritical note on the limitations of the publicly available CheXpert dataset\nprimarily on account of disparity in class distribution in training vs. testing\nsets, and unavailability of sufficient samples for few classes, which hampers\nquantitative reporting due to sample insufficiency.\n