SegAnyPET: Universal Promptable Segmentation from Positron Emission Tomography Images

• DOI: 10.48550/arxiv.2502.14351
• Publication Date: 2025-02-20
• Authors: Yichi Zhang; Le Xue; Wenbo Zhang; Lanlan Li; Yu‐Chen Liu; Chen Jiang; Yuan Cheng; Qi Yuan
• Keywords: Positron emission

Positron Emission Tomography (PET) imaging plays a crucial role in modern medical diagnostics by revealing the metabolic processes within patient's body, which is essential for quantification of therapy response and monitoring treatment progress. However, segmentation PET images presents unique challenges due to their lower contrast less distinct boundaries compared other structural modalities. Recent developments foundation models have shown superior versatility across diverse natural image tasks. Despite efforts adaptations, these works primarily focus on with detailed physiological information exhibit poor generalization ability when adapted molecular imaging. In this paper, we collect construct PETS-5k, largest dataset date, comprising 5,731 three-dimensional whole-body encompassing over 1.3M 2D images. Based established dataset, develop SegAnyPET, modality-specific 3D model universal promptable from To issue challenge discrepant annotation quality images, adopt cross prompting confident learning (CPCL) strategy an uncertainty-guided self-rectification process robustly learn high-quality labeled data low-quality noisy data. Experimental results demonstrate that SegAnyPET can correctly segment seen unseen targets using only one or few prompt points, outperforming state-of-the-art task-specific fully supervised higher accuracy strong segmentation. As first believe will advance applications various downstream tasks