Resetting the baseline: CT-based COVID-19 diagnosis with Deep Transfer Learning is not as accurate as widely thought

Abstract

Deep learning is gaining instant popularity in computer aided diagnosis of\nCOVID-19. Due to the high sensitivity of Computed Tomography (CT) to this\ndisease, CT-based COVID-19 detection with visual models is currently at the\nforefront of medical imaging research. Outcomes published in this direction are\nfrequently claiming highly accurate detection under deep transfer learning.\nThis is leading medical technologists to believe that deep transfer learning is\nthe mainstream solution for the problem. However, our critical analysis of the\nliterature reveals an alarming performance disparity between different\npublished results. Hence, we conduct a systematic thorough investigation to\nanalyze the effectiveness of deep transfer learning for COVID-19 detection with\nCT images. Exploring 14 state-of-the-art visual models with over 200 model\ntraining sessions, we conclusively establish that the published literature is\nfrequently overestimating transfer learning performance for the problem, even\nin the prestigious scientific sources. The roots of overestimation trace back\nto inappropriate data curation. We also provide case studies that consider more\nrealistic scenarios, and establish transparent baselines for the problem. We\nhope that our reproducible investigation will help in curbing hype-driven\nclaims for the critical problem of COVID-19 diagnosis, and pave the way for a\nmore transparent performance evaluation of techniques for CT-based COVID-19\ndetection.\n

Ähnliche Arbeiten