Automatic detection, segmentation, and classification of primary bone tumors and bone infections using an ensemble...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Accurate differentiation between primary bone tumors (PBTs) and bone infections is crucial for effective patient management.
• Multi-parametric Magnetic Resonance Imaging (MRI) offers rich information but requires expert interpretation for complex cases.

Purpose of the Study:

• To develop and validate an ensemble multi-task deep learning (DL) framework for simultaneous detection, segmentation, and classification of PBTs and bone infections using multi-parametric MRI.
• To assess the framework's performance against radiologist interpretations.

Main Methods:

• A retrospective study involving 749 patients with PBTs or bone infections from two centers.
• An ensemble DL framework integrating T1-weighted images (T1WI), T2-weighted images (T2WI), and clinical data for binary and three-category classification.
• Performance evaluation using Intersection over Union (IoU), Dice scores, and Area Under the Curve (AUC) on an external validation set.

Main Results:

• The framework achieved high detection and segmentation performance with IoUs of 0.71 ± 0.25 and Dice scores of 0.75 ± 0.26 on the external validation set.
• Achieved AUCs of 0.959 for binary and 0.900 for three-category classification, with accuracies of 90.6% and 78.3%, respectively.
• The three-category classification performance was superior to junior radiologists and comparable to senior radiologists.

Conclusions:

• The developed MRI-based ensemble multi-task DL framework demonstrates significant potential for automated and simultaneous detection, segmentation, and classification of PBTs and bone infections.
• The framework's performance surpasses that of junior radiologists, offering a valuable tool for clinical differential diagnosis and supporting treatment decisions.