Content-based medical image retrieval with opponent class adaptive margin loss

The increasing utilization of medical imaging technology with digital storage capabilities has facilitated the compilation large-scale data repositories. Fast access to image samples similar appearance suspected cases in these repositories can help establish a consulting system for healthcare professionals, and improve diagnostic procedures while minimizing processing delays. However, manual querying large is labor intensive. Content-based retrieval (CBIR) offers an automated solution based on quantitative assessment similarity features latent space. Since conventional methods hand-crafted typically show poor generalization performance, learning-based CBIR have received attention recently. A common framework this domain involves classifier-guided models that are trained detect different classes. Similarity assessments then performed captured by intermediate stages models. While powerful inter-class discrimination, they suboptimally sensitive within-class differences features. An alternative instead performs task-agnostic training learn embedding space enforces representational discriminability images. Within representational-learning framework, method triplet-wise learning addresses deficiencies point-wise pair-wise characterizing relationships between traditional triplet loss separation only subset within via manually-set constant margin value, so it lead suboptimal segregation opponent classes limited performance. To address limitations, we introduce triplet-learning novel Opponent Class Adaptive Margin (OCAM) loss. maintain optimally discriminative representations, OCAM considers among all pairs utilizes adaptive value automatically selected per dataset during course iterations. performance compared against state-of-the-art functions three public databases (gastrointestinal disease, skin lesion, lung disease). On average, shows mAP 86.30% KVASIR dataset, 70.30% ISIC 2019 85.57% X-RAY dataset. Comprehensive experiments each application demonstrate superior competing at 1.52%, 2.29%, non-triplet 4.56%.