New AI Framework Improves Bone Infection Segmentation from PET-CT Scans

Early and accurate diagnosis of bone infections is critical, and PET-CT imaging combines metabolic and anatomical data. However, lesion segmentation is challenging due to indistinct boundaries and inconsistent annotations from different experts or automated systems. Existing models often force a singular consensus, losing valuable diagnostic diversity.

To address this, the team proposes a decoupled dual-source learning framework. Two parallel models are trained independently on expert annotations driven by high-sensitivity and high-specificity clinical intents. An early-fusion multimodal backbone integrates PET and CT data end-to-end. To avoid performance inflation from inter-slice correlation in small datasets, they implement rigorous patient-level 3D volumetric evaluation and cross-validation. Results show that models successfully internalize distinct expert philosophies, and the cross-evaluation matrix quantifies performance variations. This diversity-preserving approach provides a robust blueprint for deploying AI in clinical bone infection segmentation.