Progressive Reliability-Guided Refinement with Cross-Modal Consistency and Iterative Correction for Weakly Supervised PET/CT Tumor Segmentation

Accurate tumor segmentation in Positron Emission Tomography/Computed Tomography (PET/CT) is essential for clinical diagnosis, staging, and treatment planning. However, the reliance on time-consuming pixel-level annotations limits the practical deployment of deep learning-based segmentation methods. Weakly supervised segmentation using image-level classification labels provides a promising alternative, yet generating accurate supervision remains highly challenging. Specifically, PET provides critical metabolic information but is susceptible to physiological hypermetabolism and background noise, while CT offers anatomical structure but suffers from low soft-tissue contrast and boundary uncertainty. Consequently, static or single-modality pseudo-labels derived from Class Activation Maps (CAMs) often suffer from boundary ambiguity and false-positive activations, failing to reliably guide segmentation. To address the core problem of dynamically evaluating and correcting supervision reliability, we propose a Progressive Reliability-Guided Refinement (PRGR) framework. Specifically, pixel-level supervision reliability is modeled by integrating PET metabolic consistency and CT structural gradients, leveraging complementary cross-modal information without introducing additional trainable parameters. Based on the estimated reliability, the image domain is adaptively partitioned into regions with different supervision credibility levels, enabling progressive optimization with differentiated learning constraints. Furthermore, pseudo-labels are iteratively updated using reliability estimates and segmentation predictions to progressively correct erroneous supervision. Experiments on two public PET/CT datasets demonstrate that our method effectively suppresses false positives and improves boundary delineation. Quantitatively, the proposed PRGR achieves Dice scores of 0.673 and 0.596 on the STS and HeadNeck datasets, respectively, yielding absolute improvements of 1.8% and 8.4% over the second-best baseline. Furthermore, it achieves superior Precision scores of 0.776 and 0.684, demonstrating its effectiveness in narrowing the performance gap between weakly and fully supervised segmentation while reducing the dependence on pixel-level annotations.