Compound Figure Separation of Biomedical Images with Side Loss

Unsupervised learning algorithms (e.g., self-supervised learning, auto-encoder, contrastive learning) allow deep models to learn effective image representations from large-scale unlabeled data. In medical analysis, even unannotated data can be difficult obtain for individual labs. Fortunately, national-level efforts have been made provide efficient access biomedical previous scientific publications. For instance, NIH has launched the Open-i\(^\circledR \) search engine that provides a database with free access. However, images in publications consist of considerable amount compound figures subplots. To extract and curate subplots, many different figure separation approaches developed, especially recent advances learning. typically required resource extensive bounding box annotation train detection models. this paper, we propose simple (SimCFS) framework uses weak classification annotations images. Our technical contribution is three-fold: (1) introduce new side loss designed separation; (2) an intra-class augmentation method simulate hard cases; (3) proposed enables deployment classes images, without requiring annotations. From results, SimCFS achieved state-of-the-art performance on ImageCLEF 2016 Compound Figure Separation Database. The source code publicly available at https://github.com/hrlblab/ImageSeperation.