Dual Bounding Box for Medical Image Segmentation

Segmentation in medical imaging is crucial for accurately identifying anatomical structures and pathological areas and is essential for precise diagnosis and treatment planning. Although bounding boxes have been widely used as a weak labeling method for training segmentation models due to their annotation efficiency, we propose a novel dual bounding box approach. This method employs two bounding boxes per class: an inner box focusing on the core target and an outer box capturing peripheral regions and contextual information. Unlike traditional single-bounding box methods, the dual-box approach provides enhanced guidance to the model, enabling it to incrementally refine segmentation boundaries and address challenges such as boundary ambiguity. Importantly, this method does not require significant additional annotation effort, as both bounding boxes can be easily derived during the annotation process. By combining weak labeling efficiency with improved segmentation accuracy, our dual-bounding box approach offers a cost-effective solution for medical image segmentation, providing a practical compromise between annotation ease and clinical utility.