2020/07/04 by Issam Laradji, Laradji, Issam, Pau Rodríguez +15
Medicine · Computer Science · #COVID-19 diagnosis using AI #Radiomics and Machine Learning in Medical Imaging #AI in cancer detection
paper · pdf · doi:10.48550/arxiv.2007.02180
Coronavirus Disease 2019 (COVID-19) has spread aggressively across the world\ncausing an existential health crisis. Thus, having a system that automatically\ndetects COVID-19 in tomography (CT) images can assist in quantifying the\nseverity of the illness. Unfortunately, labelling chest CT scans requires\nsignificant domain expertise, time, and effort. We address these labelling\nchallenges by only requiring point annotations, a single pixel for each\ninfected region on a CT image. This labeling scheme allows annotators to label\na pixel in a likely infected region, only taking 1-3 seconds, as opposed to\n10-15 seconds to segment a region. Conventionally, segmentation models train on\npoint-level annotations using the cross-entropy loss function on these labels.\nHowever, these models often suffer from low precision. Thus, we propose a\nconsistency-based (CB) loss function that encourages the output predictions to\nbe consistent with spatial transformations of the input images. The experiments\non 3 open-source COVID-19 datasets show that this loss function yields\nsignificant improvement over conventional point-level loss functions and almost\nmatches the performance of models trained with full supervision with much less\nhuman effort. Code is available at:\n urlhttps://github.com/IssamLaradji/covid19weaksupervision.\n