Foreground Background Difference Knowledge-Based Small Sample Target Segmentation for Image-Guided Radiation Therapy

Abstract

The aim of this paper is to exploit a small sample (data scarcity) target segmentation technique for image-guided radiation therapy. The technique is grounded on a prototype-based approach—widely used small sample segmentation method. In this paper, we propose a foreground–background difference knowledge learning framework to perform the small sample target segmentation task. Its main differences from the traditional prototype-based approaches and novel contributions may be enumerated in two aspects: (1) A subdivision strategy to generate multiple foreground–background prototypes for each class in the support images, and the generated prototype is used to build a collection of query foreground and background prototypes. (2) A cross-prototype attention module to learn the correlation and difference knowledge of inter-class prototypes and transfer the knowledge to the query prototype for iterative updates. The main advantage of our framework is that: (1) the intra-class prototype set can comprehensively reflect the class features, avoiding the high computational complexity caused by dense matching; and (2) knowledge of inter-class differences provides comprehensive foreground–background segmentation information, greatly supporting accurate segmentation of the query set. In the 5-shot SegRap dataset experiment, the proposed model achieved Dice coefficients of 82.23% in the same-domain setting and 81.01% in the cross-domain setting. Similarly, in the 5-shot HECKTOR2022 dataset experiment, it achieved 83.59% in the same-domain setting and 81.48% in the cross-domain setting. For the 5-shot BTCV and CHAOS datasets, the model attained Dice coefficients of 79.00% and 79.70%, respectively. These results demonstrate the model's accuracy, efficiency, and generalization. This study presents a significant advancement in medical image segmentation by introducing a prototype-based model that effectively addresses data scarcity. By leveraging intra- and inter-class attention mechanisms, the model ensures robust generalization and reliable performance across datasets, paving the way for efficient and precise clinical applications with minimal reliance on large annotated datasets.