Efficient management of medical image datasets for content-based case retrieval via sparse annotation and bidirectional label propagation

Abstract

Efficient management of medical image datasets is critical for clinical decision-making. However, current methods lack fine-grained annotation management and content-based case retrieval capabilities. Existing systems struggle to handle the annotation burden and to efficiently retrieve similar cases based on image content. To address these challenges, we propose a novel method that uses sparse annotation at the beginning, middle, and end of key disease regions, thereby reducing annotation effort while maintaining segmentation accuracy. We introduce CA-Morpher, an unsupervised image registration model enhanced by a cross-attention mechanism that effectively propagates sparse labels. Additionally, we develop a bidirectional label transfer algorithm (BLTA) that combines prior annotations and registration to propagate labels through bidirectional transfer and pseudo-label weighted fusion. This approach improves the management of annotated datasets by enabling content-based case retrieval, thereby enhancing overall data management efficiency. Experimental results show that our method achieves a Dice score of 76.62%, a Jaccard index of 63.25%, and a Hausdorff distance (HD) of 4.64 on the medical segmentation decathlon (MSD) pancreatic tumor dataset, outperforming current sparse annotation methods and significantly improving the management and retrieval of medical image data.