Cross-Source Supervision for Bone Infection Segmentation in Dual-Modality PET-CT

2026-05-19 · Source: cs.CV updates on arXiv.org · Field: Health & Wellbeing — Medical Specialties & Subspecialties, Medical Devices & Health Technology, Artificial Intelligence & Machine Learning · Depth: Expert, long

Summary

A new bimodal end-to-end segmentation framework has been developed for bone infection segmentation in dual-modality PET-CT scans. This framework addresses challenges like indistinct lesion boundaries and annotation inconsistencies by integrating PET metabolic signals and CT bone-window anatomy through an early-fusion multimodal representation. To counter performance inflation from inter-slice correlation in small datasets, the study employs a rigorous patient-level 3D volumetric evaluation and cross-validation, discarding traditional 2D methods. It also introduces a decoupled dual-source learning framework, training parallel models on independent expert annotations driven by high-sensitivity and high-specificity clinical intents. Experimental results demonstrate the effectiveness of multimodal PET-CT fusion and reveal how models internalize distinct expert diagnostic philosophies, providing a robust, diversity-preserving paradigm for clinical AI deployment.

Key takeaway

For AI Scientists developing medical imaging segmentation models, you should consider adopting a decoupled dual-source learning approach when faced with significant inter-observer variability in ground truth annotations. This strategy allows your models to internalize distinct clinical philosophies (e.g., high-sensitivity vs. high-specificity), leading to more robust and clinically relevant outputs than forcing a single consensus. Ensure rigorous 3D patient-level evaluation to validate performance, especially with limited datasets.

Key insights

Decoupled dual-source learning with PET-CT fusion improves bone infection segmentation by accommodating diverse clinical annotation philosophies.

Principles

• Multimodal fusion enhances disease detection.
• Patient-level 3D evaluation mitigates small dataset bias.
• Annotation variability can be modeled, not just averaged.

Method

The method uses an early-fusion dual-channel U-Net, trained with a decoupled dual-source learning framework on independent high-sensitivity and high-specificity annotations, and evaluated via patient-level 3D volumetric cross-evaluation.

In practice

• Integrate PET and CT for improved bone infection diagnosis.
• Train separate models for high-sensitivity screening and high-specificity surgical guidance.
• Use 3D volumetric evaluation for small medical image datasets.

Topics

• Osteomyelitis
• PET-CT
• Multimodal Segmentation
• Annotation Variability
• Dual-Source Learning

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Mutual Distillation of Dual-Foundation Models for Semi-Supervised PET/CT Segmentation — Mutual distillation of dual-foundation models improves semi-supervised PET/CT segmentation with minimal labeled data.
• Dual-Modal Lung Cancer AI: Interpretable Radiology and Microscopy with Clinical Risk Integration — Integrating CT radiology and H&E histopathology via dual-modal AI improves lung cancer diagnosis and interpretability.
• Dual-Domain Equivariant Generative Adversarial Network for Multimodal CT-PET Synthesis — Dual-domain learning and rotational equivariance significantly enhance multimodal medical image synthesis accuracy and robustness.
• Intracranial Aneurysm Classification and Segmentation via Tri-Axial ROI and Multi-Task Learning — A multi-task framework enhances intracranial aneurysm analysis by integrating classification and multi-class segmentation across diverse modalities.
• Align then Refine: Text-Guided 3D Prostate Lesion Segmentation — Text-guided, phase-scheduled training improves 3D prostate lesion segmentation by integrating semantic alignment and confidence-aware refinement.
• Supervise-assisted Multi-modality Fusion Diffusion Model for PET Restoration — MFdiff restores high-quality PET from low-dose inputs using multi-modality fusion and a two-stage supervised diffusion model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.