AdaTriplet: Adaptive Gradient Triplet Loss with Automatic Margin Learning for Forensic Medical Image Matching

This paper tackles the challenge of forensic medical image matching (FMIM) using deep neural networks (DNNs). FMIM is a particular case content-based retrieval (CBIR). The main in compared to general CBIR, that subject whom query belongs may be affected by aging and progressive degenerative disorders, making it difficult match data on level. CBIR with DNNs generally solved minimizing ranking loss, such as Triplet loss (TL), computed representations extracted DNN from original data. TL, particular, operates triplets: anchor, positive (similar anchor) negative (dissimilar anchor). Although TL has been shown perform well many tasks, still limitations, which we identify analyze this work. In paper, introduce (i) AdaTriplet – an extension whose gradients adapt different difficulty levels samples, (ii) AutoMargin method technique adjust hyperparameters margin-based losses our proposed dynamically. Our results are evaluated two large-scale benchmarks for based Osteoarthritis Initiative Chest X-ray-14 datasets. codes allowing replication study have made publicly available at https://github.com/Oulu-IMEDS/AdaTriplet .