Bone Infection Segmentation in PET-CT Using Cross-Source Supervision
A new study tackles the challenge of segmenting bone infections in PET-CT scans, where annotation discrepancies between experts and automated systems complicate accurate lesion delineation. The researchers propose a bimodal end-to-end segmentation framework that fuses PET metabolic signals with CT bone-window anatomy via early-fusion multimodal representation. To avoid performance inflation from inter-slice correlation in small datasets, they replace traditional 2D evaluation with rigorous patient-level 3D volumetric assessment. The work is published on arXiv under reference 2605.16373.
Key facts
- PET-CT integrates anatomical CT and metabolic PET for bone infection diagnosis.
- Lesion segmentation is hindered by indistinct boundaries and annotation inconsistencies.
- A bimodal end-to-end segmentation framework is developed using early-fusion multimodal representation.
- The study discards 2D evaluation methods in favor of patient-level 3D volumetric evaluation.
- The research is published on arXiv with ID 2605.16373.
- The framework integrates PET metabolic signals and CT bone-window anatomy.
- The approach addresses performance inflation from inter-slice correlation in small datasets.
- The work aims to improve early and accurate diagnosis and lesion localization.
Entities
Institutions
- arXiv