The Value of Visual Attention for COVID-19 Classification in CT Scans

Adrit Rao, Jongchan Park, Oliver Aalami; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 433-438

Detecting COVID-19 in early stages is crucial in order to initiate timely treatment of disease. COVID-19 screening with chest CT scans has been utilized due to the rapidity of results and robustness. Computer vision aided medical diagnosis with deep learning models can improve accuracy and efficiency of screening. When developing models for high-risk medical classification tasks, it is important to aim to reach radiologist level interpretation in terms of cognition. When the human brain analyzes visual information, cognitive visual attention is applied in order to apply more focus onto higher frequency regions of interest. Using attention mechanisms in order to infer channel and spatial attention maps within convolutional neural networks can improve the performance in classification of COVID-19 changes. Through performing a compact study with a quantitative accuracy measure along with a qualitative visualization of activation heat-maps, we study the benefits of visual self-attention for the classification of COVID-19.