BoneCoT: multicentre validation of a whole-body skeleton foundation model for bone metastases guided by clinician-derived chain of thought

2026-07-02 · Source: Machine learning : nature.com subject feeds · Field: Health & Wellbeing — Artificial Intelligence & Machine Learning, Health & Medical Research, Clinical Care & Medical Practice · Depth: Expert, long

Summary

BoneCoT is a novel whole-body skeleton foundation model designed for detecting bone metastases using computed tomography (CT) images, enhanced by a clinician-derived chain-of-thought (CoT) fine-tuning approach. The model was pretrained on 29.3 million CT images from 30,267 patients across 12 skeletal sites. It was then refined over a graph of 26 clinically relevant tasks, including diagnosis, complications, tumor type, and biomarkers. In evaluations across these 26 tasks and multicentre cohorts from 10 hospitals, BoneCoT significantly outperformed existing advanced methods by 20% in area under the receiver operating characteristic curve (AUROC). Notably, BoneCoT achieved a 40% AUROC improvement in distinguishing primary from metastatic lesions, surpassing experienced radiologists. This model aims to provide expert-level insights by integrating multidisciplinary information for precise bone metastasis diagnosis.

Key takeaway

For radiologists and oncologists evaluating bone metastases, BoneCoT offers a validated AI tool that significantly improves diagnostic precision, especially in differentiating primary from metastatic lesions. You should consider integrating such clinician-guided foundation models into your diagnostic workflow to augment accuracy and reduce discrepancies. This technology could streamline multidisciplinary collaboration and enhance patient care by providing expert-level insights from CT imaging.

Key insights

Clinician-derived reasoning can significantly enhance AI models for complex medical diagnostic tasks like bone metastasis detection.

Principles

• Integrating multidisciplinary clinical reasoning improves AI diagnostic accuracy.
• Foundation models benefit from task-specific fine-tuning on diverse clinical data.
• AI can surpass human expert performance in specific, complex diagnostic differentiations.

Method

BoneCoT pretrains on 29.3 million CT images, then fine-tunes using a chain-of-thought approach over a graph of 26 clinical tasks, incorporating task interdependencies for multimodal reasoning.

In practice

• Implement chain-of-thought fine-tuning for medical imaging AI.
• Utilize large-scale, multi-center CT datasets for robust model pretraining.
• Focus AI development on differentiating challenging lesion types.

Topics

• Bone Metastases
• Computed Tomography
• Foundation Models
• Chain-of-Thought
• Medical Imaging AI
• Diagnostic Accuracy

Code references

• FrankZhangRp/BoneCoT

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Domain Expert

Related on AIssential

• Learning neuroimaging models from health system-scale data — Health system-scale data enables AI foundation models to achieve superior, fair, and explainable neuroimaging diagnostics.
• Meta-encoder: a unified integration framework for multiple pathological foundation models in cancer detection — The Meta-Encoder integrates multiple pathological foundation models to improve cancer detection and molecular characterization.
• Cross-Source Supervision for Bone Infection Segmentation in Dual-Modality PET-CT — Decoupled dual-source learning with PET-CT fusion improves bone infection segmentation by accommodating diverse clinical annotation philosophies.
• CORTEX: A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs — CORTEX provides structured, verifiable reasoning traces for 3D chest CT MLLMs, mirroring clinical diagnostic workflows.
• Learning Emergent Modular Representations in Multi-modality Medical Vision Foundation Models — DEX uses modular networks with experts and a director to manage modality-specific and shared representations in multi-modality medical FMs.
• Phenotyping TPF via Self-Supervised Learning: A Label-Agnostic Framework with Expert Validation — Self-supervised learning can discover stable, clinically interpretable fracture phenotypes without relying on inconsistent human labels.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.