Transformer-based multi-scale feature fusion for real-time CT bone metastasis detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Bone metastasis is a common, serious complication of advanced cancers.
• Early and accurate detection of bone metastatic lesions is crucial for timely intervention and improved patient outcomes.
• Computed tomography (CT) is vital for non-invasive bone lesion identification, but challenges remain in detecting small, low-contrast lesions with deep learning models.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient detection of bone metastases from CT scans.
• To address the limitations of existing models in handling low-contrast lesions and complex tumor microenvironments.
• To enhance the real-time applicability and accuracy of AI in diagnosing bone metastatic disease.

Main Methods:

• Proposed BM-DETR, a Transformer-based model incorporating spatial-contextual enhancement module (SCEM), AttentionUpsample, and dilated transformer attention block (DTAB).
• SCEM enhances low-contrast lesion features via channel attention and spatial mixing.
• AttentionUpsample fuses multi-scale features, while DTAB improves contextual modeling and efficiency.

Main Results:

• BM-DETR achieved a mean Average Precision (mAP50) of 0.9376 on the OsteoScan dataset.
• BM-DETR achieved a mAP50 of 0.9139 on the BMSeg dataset, outperforming state-of-the-art methods.
• The model demonstrated a balance between high accuracy and computational efficiency.

Conclusions:

• BM-DETR offers a robust solution for automated bone metastasis detection, significantly advancing diagnostic capabilities.
• The model's efficiency and accuracy support potential edge deployment for early screening and intelligent diagnosis.
• This work lays a foundation for clinical translation of AI-driven diagnostic systems for bone metastases.