Automated diagnosis of cancer bone metastases using multi-view contrastive learning

Abstract Background. Functional imaging, particularly Single Photon Emission Computed Tomography (SPECT) bone scintigraphy, plays a critical role in diagnosing and assessing various diseases such as bone metastases. Deep learning has shown promise as a robust tool for automated disease diagnosis by extracting meaningful features from medical images. However, challenges such as limited sample availability, costly annotation, and severe class imbalance often compromise model performance, representing a persistent limitation in prior research. As a result, effectively learning discriminative features from bone scintigraphy datasets with limited and imbalanced samples remains a major challenge. Objective . This study introduces a novel Multi-view Contrastive Learning approach to enhance the automated detection of bone metastases and differentiate between bone metastases originating from different solid cancers. Methods. We propose a multi-class classification model based on multi-view contrastive learning, implemented within a transfer learning framework comprising a two-stage process. In the upstream task, the model initializes feature representations by transferring self-supervised learning weights pretrained on a general large-scale dataset. Subsequently, a contrastive learning network extracts deep feature representations from external data sources. In the downstream task, these learned representations are fine-tuned on the target dataset. Furthermore, we introduce a fusion module to integrate multi-view features, improving diagnostic performance under multi-view conditions. Results. The proposed model achieved an accuracy of 0.9195 in the binary classification task and 0.8725 in the three-class classification task, which includes differentiation of primary cancer types—representing a 6.25 percentage point improvement over traditional supervised methods. Notably, our framework demonstrated accuracies of 0.6763 and 0.6645 in distinguishing bone metastases from lung and breast cancer, respectively, providing the first exploratory evidence of primary cancer inference from bone scans. These results validate the effectiveness of contrastive pretraining in addressing class imbalance and data scarcity and demonstrate the superiority of the proposed method in multi-class classification tasks. Conclusion. Our method achieved competitive performance in diagnosing bone metastases using whole-body SPECT imaging, demonstrating its effectiveness in extracting discriminative features from limited and imbalanced bone scintigraphy datasets. Additionally, we conducted an initial exploration into classifying bone metastases according to their primary cancer origin, specifically distinguishing between lung and breast cancer metastases.