DBFANet: a dual-branch feature alignment network for automated detection of breast cancer bone metastasis.

Bone scan imaging for the detection of bone metastasis of breast cancer has been widely adopted; however, noise, anatomy superimposition, and small size for early lesions will severely affect its prediction performance. In this work, we propose a new framework with two major contributions to solve the main problems existing in current deep-learning-based approaches.In this study, we put forward a new model called the dual branch feature alignment network (DBFANet) for automated breast cancer bone metastases detection in bone scintigraphy. DBFA-net adopts a dual-branch CNN-Transformer structure: the CNN branch focuses on the local details, while the Transformer branch learns the global context. In addition, we design a feature alignment module (FRAT), which employs the bi-directional cross-attention mechanism for the complementary feature from two branches. Moreover, we propose an enhanced multi-scale attention module (EMSA) based on the squeeze-and-excitation (SE) block for stronger multi-scale lesion representations with less background noise suppression.We validated our proposed model based on a bone scintigraphy dataset containing 5092 images. In terms of bone metastasis prediction, DBFANet achieved an accuracy, precision, and recall value of 93.1%, 84.6%, and 84.7%, respectively, all superior to previous models (such as ResNet-50, EfficientNet-V2, and MaxViT). The ablation study has shown that both FRAT and EMSA have individual effectiveness and complementary benefits. Finally, additional external validation was performed on a publicly available bone scintigraphy dataset (BS-80K).DBFANet shows the highest detection performance for bone metastasis detection from multiview bone scintigraphy images with imbalanced classes and noise in the image, and the feature alignment with enhanced multiscale attention of DBFANet provides a useful and precise tool for bone metastasis diagnosis in a nuclear medicine imaging scenario.