From Slice to Volume: Weakly-Supervised Learning with RECIST Annotations for Lesion Segmentation

Accurate tumor size measurement is essential for monitoring disease progression and evaluating therapeutic efficacy. While manual lesion volume segmentation remains the clinical standard, this process is both labor intensive and highly dependent on specialized radiological expertise. The widespread availability of RECIST annotations present an opportunity to develop automated lesion segmentation solutions. This paper proposes a novel weakly supervised segmentation pipeline that effectively leverages RECIST annotations through a geometric trimapbased mask generation method for RECIST slices, alongside a framework designed for non-RECIST slices. This framework integrates two synergistic modules: (1) a registration module that leverages the inherent spatial consistency of medical volumes, and (2) a segmentation module incorporating co-regularization mechanisms to improve robustness against pseudo-label noise. These modules interactively refine each other to enhance segmentation accuracy. Comprehensive evaluations demonstrate the superiority of our framework in the task of RECIST weakly supervised volumetric segmentation.