Chest X-Ray Feature Pyramid Sum Model with Diseased Area Data Augmentation Method

Le résumé fourni par la source

Deep learning has shown considerable promise in medical image analysis, but significant challenges remain. These stem from the inherent complexities of medical images, such as varying sizes of lesions within the same image and the potential coexistence of multiple diseases. To address these issues, we propose a novel model combining TResNet with Feature Pyramid Network (FPN). This model adeptly handles multi-label classification, demonstrating robust performance across a range of lesion sizes. Furthermore, most medical images follow a long-tail distribution, presenting class imbalance problems, where the occurrence of one lesion often correlates with the presence of others. Considering these correlations, we introduced a strategy for dealing with the class imbalance issue by augmenting minority classes using bounding box information of the disease. Our proposed approach offers a novel solution for handling the unique challenges in deep learning-based medical image analysis, paving the way for more precise interpretations of complex medical images. The performance of mAP in 26 disease classes has been improved from 32.76% to 33.37% in a single model, and 35.11% in ensemble model.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.