Automated Detection and Segmentation of Bone Metastases on Spine MRI Using U-Net: A Multicenter Study

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Spinal bone metastases are a common complication of cancer.
• Accurate segmentation and detection on MRI are crucial for treatment planning.
• Manual analysis of spinal MRI for bone metastases is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and evaluate deep learning models for automated segmentation and detection of bone metastases on spinal MRI.
• To compare the performance of different U-Net architectures and MRI sequences.
• To assess the diagnostic performance against experienced radiologists.

Main Methods:

• Whole spine MRI scans from 322 adult patients with bone metastasis were used.
• Seven 2D and 3D U-Net models were trained using combinations of T1-weighted, contrast-enhanced T1-weighted Dixon fat-only, and contrast-enhanced fat-suppressed T1-weighted sequences.
• Performance was evaluated using Dice coefficient, pixel-wise recall/precision, and per-lesion sensitivity, with comparison to radiologist performance on an external test set.

Main Results:

• The 2D U-Net T1 + CE model achieved the best segmentation performance (Dice coefficient 0.699, recall 0.653) on the external test set.
• This model demonstrated high per-lesion sensitivity for metastasis detection (0.828 internal, 0.857 external).
• Radiologists achieved a mean per-lesion sensitivity of 0.746 and positive predictive value of 0.701 on the external test set.

Conclusions:

• Deep learning models, particularly the 2D U-Net T1 + CE model, show high diagnostic performance for automated bone metastasis detection on spinal MRI.
• These models offer a promising tool to aid radiologists in the accurate and efficient identification of spinal bone metastases.
• Automated analysis has the potential to improve consistency and reduce workload in the interpretation of spinal MRI for metastatic disease.