FAMSA-Net: frequency-aware multi-scale spatial attention network for gain quality grading in automated breast ultrasound images

Gain setting affects Automated Breast Ultrasound (ABUS) image brightness and contrast, inappropriate gain may obscure lesion boundaries, alter tissue echogenicity, and degrade clinical interpretation. Current gain adjustment relies heavily on operator experience, resulting in intra- and inter-operator variability. This study aims to develop an automated and objective gain quality grading framework for ABUS images to reduce subjective variability and facilitate standardized image acquisition. A large-scale ABUS dataset comprising 31,628 images was collected from Yunnan Cancer Hospital, including five gain quality categories: Insufficient, Low, Optimal, High, and Excessive. Frequency-aware Multi-scale Spatial Attention Network (FAMSA-Net) was proposed for gain quality grading. Gain-discriminative residual blocks and a Multi-Scale Gain Fusion (MSGF) module were designed to capture gain-sensitive characteristics across multiple feature scales. In addition, we integrate Frequency-Domain Enhancement (FDE) to amplify gain-related frequency cues, Spatial Transformer Networks (STN) to spatially align breast anatomical structures, and Context-Aware Pooling (CAP) to adaptively aggregate contextual information. Experimental results demonstrated that FAMSA-Net achieved an accuracy of 95.49%, outperforming DenseNet121 (91.30%) and EfficientNetB3 (89.93%) by 4.19% and 5.56%. FAMSA-Net uses 16.58 M parameters and 4.34 GFLOPs, which is more efficient than DenseNet121 and EfficientNetB3. Ablation experiments showed that removing STN caused the largest accuracy drop, from 95.49% to 83.76%. In contrast, under the GAP pooling setting, removing FDE or MSGF improved accuracy to 90.77% and 90.01%, suggesting that the module contributions are not additive. Furthermore, the full model reached its peak accuracy of 95.49% after replacing GAP with CAP, indicating the importance of contextual aggregation for effective module cooperation.