2021/02/21 by Yen Nhi Truong Vu, Richard Wang, Vu, Yen Nhi Truong +9 · 1 citation
Computer Science · Medicine · #AI in cancer detection #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Image and Video Processing (eess.IV) #Lung Cancer Diagnosis and Treatment #Machine Learning (cs.LG) #Mycobacterium research and diagnosis #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2102.10663
openalex publication_date 2021/02/21 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Self-supervised contrastive learning between pairs of multiple views of the\nsame image has been shown to successfully leverage unlabeled data to produce\nmeaningful visual representations for both natural and medical images. However,\nthere has been limited work on determining how to select pairs for medical\nimages, where availability of patient metadata can be leveraged to improve\nrepresentations. In this work, we develop a method to select positive pairs\ncoming from views of possibly different images through the use of patient\nmetadata. We compare strategies for selecting positive pairs for chest X-ray\ninterpretation including requiring them to be from the same patient, imaging\nstudy or laterality. We evaluate downstream task performance by fine-tuning the\nlinear layer on 1% of the labeled dataset for pleural effusion classification.\nOur best performing positive pair selection strategy, which involves using\nimages from the same patient from the same study across all lateralities,\nachieves a performance increase of 14.4% in mean AUC from the ImageNet\npretrained baseline. Our controlled experiments show that the keys to improving\ndownstream performance on disease classification are (1) using patient metadata\nto appropriately create positive pairs from different images with the same\nunderlying pathologies, and (2) maximizing the number of different images used\nin query pairing. In addition, we explore leveraging patient metadata to select\nhard negative pairs for contrastive learning, but do not find improvement over\nbaselines that do not use metadata. Our method is broadly applicable to medical\nimage interpretation and allows flexibility for incorporating medical insights\nin choosing pairs for contrastive learning.\n