CT-based COVID-19 Triage: Deep Multitask Learning Improves Joint\n Identification and Severity Quantification

Abstract

The current COVID-19 pandemic overloads healthcare systems, including\nradiology departments. Though several deep learning approaches were developed\nto assist in CT analysis, nobody considered study triage directly as a computer\nscience problem. We describe two basic setups: Identification of COVID-19 to\nprioritize studies of potentially infected patients to isolate them as early as\npossible; Severity quantification to highlight studies of severe patients and\ndirect them to a hospital or provide emergency medical care. We formalize these\ntasks as binary classification and estimation of affected lung percentage.\nThough similar problems were well-studied separately, we show that existing\nmethods provide reasonable quality only for one of these setups. We employ a\nmultitask approach to consolidate both triage approaches and propose a\nconvolutional neural network to combine all available labels within a single\nmodel. In contrast with the most popular multitask approaches, we add\nclassification layers to the most spatially detailed upper part of U-Net\ninstead of the bottom, less detailed latent representation. We train our model\non approximately 2000 publicly available CT studies and test it with a\ncarefully designed set consisting of 32 COVID-19 studies, 30 cases with\nbacterial pneumonia, 31 healthy patients, and 30 patients with other lung\npathologies to emulate a typical patient flow in an out-patient hospital. The\nproposed multitask model outperforms the latent-based one and achieves ROC AUC\nscores ranging from 0.87+-01 (bacterial pneumonia) to 0.97+-01 (healthy\ncontrols) for Identification of COVID-19 and 0.97+-01 Spearman Correlation for\nSeverity quantification. We release all the code and create a public\nleaderboard, where other community members can test their models on our test\ndataset.\n

Ähnliche Arbeiten