FGM-SPCL:Open-Set Recognition Network for Medical Images Based on Fine-Grained Data Mixture and Spatial Position Constraint Loss  

作　　者：Ruru ZHANG Haihong E Lifei YUAN Yanhui WANG Lifei WANG Meina SONG 

机构地区：[1]School of Computer Science,Beijing University of Posts and Telecommunications,Beijing 100876,China [2]Engineering Research Center of Information Networks,Ministry of Education,Beijing 100876,China [3]Hebei Eye Hospital,Xingtai 054001,China

出　　处：《Chinese Journal of Electronics》2024年第4期1023-1033,共11页电子学报（英文版）

基　　金：National Natural Science Foundation of China (Grant No. 62176026);Beijing Natural Science Foundation (Grant No. M22009);Key Research and Development Program of Hebei (Grant No. 22377775D);Engineering Research Center of Information Networks, Ministry of Education。

摘　　要：The current intelligent auxiliary diagnosis models all follow the closed-set recognition setting. After the model is deployed online, the input data is often not completely controlled. Diagnosing an untrained disease as a known category would lead to serious medical malpractice. Therefore, realizing the open-set recognition is significant to the safe operation of the intelligent auxiliary diagnosis model. Currently, most open-set recognition models are studied for natural images, and it is very challenging to obtain clear and concise decision boundaries between known and unknown classes when applied to fine-grained medical images. We propose an open-set recognition network for medical images based on fine-grained data mixture and spatial position constraint loss(FGM-SPCL) in this work.Considering the fine graininess of medical images and the diversity of unknown samples, we propose a fine-grained data mixture(FGM) method to simulate unknown data by performing a mixing operation on known data to expand the coverage of unknown data difficulty levels. In order to obtain a concise and clear decision boundary, we propose a spatial position constraint loss(SPCL) to control the position distribution of prototypes and samples in the feature space and maximize the distance between known classes and unknown classes. We validate on a private ophthalmic OCT dataset, and extensive experiments and analyses demonstrate that FGM-SPCL outperforms state-of-the-art models.

关 键 词：Few-shot class-incremental learning Embedding augmentation Classifier adaptation Image classification 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] R319[医药卫生—基础医学]