Researchers have demonstrated that an attention-gated multi-task deep-learning model can jointly classify and segment breast ultrasound lesions with performance statistically comparable to dedicated single-task networks while cutting parameter requirements by 46 percent. On a held-out test set of 117 images from the publicly available BUSI dataset, the unified model achieved a classification accuracy of 93.16 percent, a macro F1-score of 0.931, and a macro area under the receiver operating characteristic curve of 0.984. These classification outcomes were statistically comparable to a dedicated classification-only baseline that recorded an accuracy of 91.45 percent, yielding a paired McNemar test p-value of 0.63.
For lesion segmentation, the unified architecture achieved a mean Dice similarity coefficient of 0.766, with a bootstrap 95 percent confidence interval spanning 0.710 to 0.819 across 1,000 resamples. This segmentation performance proved statistically comparable to the dedicated segmentation-only baseline’s Dice score of 0.810 (paired Wilcoxon signed-rank test p = 0.22). Incorporating attention gates into the segmentation-only baseline did not produce a statistically significant change in segmentation accuracy, resulting in a Dice score of 0.813 versus 0.810 (p = 0.50). Furthermore, the unified network demonstrated 100 percent recall for normal scans and misclassified no malignant cases as normal.
Built upon a shared ResNet-34 encoder, the multi-task framework simultaneously executes three-class classification across normal, benign, and malignant cases alongside lesion segmentation within a single forward pass. The researchers developed and evaluated the architecture using the 780-image BUSI dataset split into stratified subsets of 70 percent for training, 15 percent for validation, and 15 percent for testing using a fixed random seed for reproducibility. The unified model was compared directly against three dedicated baselines: a classification-only model, a segmentation-only model, and an attention-gated segmentation-only model.
Breast ultrasound is widely deployed for breast cancer screening, but its clinical interpretation is operator-dependent, and conventional computer-aided diagnosis pipelines address classification and segmentation through separate models that elevate parameter budgets and deployment complexity. By consolidating these tasks into a single architecture, the unified model operated with 24.7 million parameters compared to the 45.9 million parameters required by two dedicated models. The authors concluded that the unified attention-gated framework provides practical value through parameter efficiency and workflow consolidation rather than through accuracy gains over dedicated single-task approaches.
Subject of Research: Unified multi-task deep learning for breast ultrasound classification and segmentation
Article Title: A unified attention-gated multi-task network for joint classification and lesion segmentation of breast ultrasound images
Article References: Zhang, Y., Liang, S.-M., Liu, J., Wen, W., & Fan, Y.-J. (2026). A unified attention-gated multi-task network for joint classification and lesion segmentation of breast ultrasound images. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-03007-6
Image Credits: AI Generated
DOI: 10.1186/s12880-026-03007-6
Keywords: breast ultrasound, deep learning, multi-task learning, lesion segmentation, computer-aided diagnosis, ResNet-34, BUSI dataset, unified, attention-gated, multi-task, network, joint
News Source: Ophelia Keating. (November 5, 2026). Multi-Task AI Matches Separate Models in Breast Ultrasound Analysis. Scienmag.



