Attention-driven multi-sequence MRI representations for breast cancer diagnosis

Magnetic resonance imaging (MRI) is the recommended imaging modality for breast cancer diagnosis; however, classifying benign and malignant breast lesions using multi-sequence MRI remains a significant challenge. This is primarily due to the heterogeneity and complexity of breast lesions, along with the substantial imaging data provided by each sequence of MRI. These factors place high demands on the expertise of clinicians, and the processing of extensive MRI scans is both costly and prone to error. Different sequence MRI reveals diverse characteristics of the lesions, and joint analysis of multi-sequence data has greater diagnostic value for breast cancer. In recent years, some studies have attempted to employ deep learning methods for multi-sequence MRI fusion to enhance diagnostic performance. However, existing approaches often lack robust strategies for feature learning that are specifically tailored to the distinct characteristics of each sequence MRI. Additionally, they have not fully leveraged the relevant information from multi-sequence MRI during modeling, limiting the effectiveness of the model. Inspired by the diagnostic workflow of radiologists, an Attention-driven Feature Learning and Fusion (AFLF) framework was proposed to classify benign and malignant breast lesions using multi-sequence MRI. Our framework employs an attention-based encoding network to learn attention-aware representations for each sequence MRI, enabling the model to focus on the unique lesion characteristics in each sequence. Furthermore, an adaptive attention learning fusion module facilitates the interaction and fusion of these representations, ensuring the relevance and representativeness of the features from different sequences are fully leveraged to achieve efficient breast cancer diagnosis. Experiments on MRI scans from 318 patients demonstrate that the AFLF outperforms existing state-of-the-art algorithms, achieving a classification accuracy of 93.75% and an AUC of 98.89%¹.