Detection of Bone Metastases on Bone Scans through Image Classification with Contrastive Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Bone metastases significantly impact patient prognosis.
• Bone scans are crucial for diagnosis but have limited accuracy due to non-specific findings.
• Deep learning offers potential to enhance diagnostic efficacy.

Purpose of the Study:

• To evaluate deep learning techniques for improving bone metastasis detection on bone scans.
• To assess the performance of various deep learning models, including CNN, ResNet, and DenseNet, with and without contrastive learning.
• To determine the clinical utility of these models in aiding diagnosis and patient management.

Main Methods:

• Retrospective analysis of 19,041 patients undergoing bone scans (May 2011 - Dec 2019).
• Development and application of binary classification deep learning algorithms (CNN, ResNet, DenseNet) on 37,427 bone scan image sets.
• Comparison of model performance with and without contrastive learning, using physician-reviewed reports as the reference standard.

Main Results:

• All evaluated deep learning models showed improved performance with contrastive learning.
• The optimal model achieved high accuracy (0.961), precision (0.878), F1 score (0.712), AUC (0.92), and negative predictive value (NPV) (0.965).
• High NPV indicates strong potential for safely ruling out bone metastases.

Conclusions:

• Deep learning, particularly with contrastive learning, enhances bone metastasis detection accuracy on bone scans.
• The high NPV of the optimal model can assist physicians in excluding metastases, reducing workload.
• Improved diagnostic accuracy can lead to better patient care and management strategies.