MRAE: Multi-Resolution Attention Ensemble with Hybrid CNN–Transformer Fusion for Breast Ultrasound Classification

Breast ultrasound images can be classified as benign, malignant, and normal. Due to the imbalanced distribution of classes in breast ultrasound images, intra-class heterogeneity of lesions, and ultrasound artifacts like speckle noise, classification of breast ultrasound images remains a challenging problem. In this paper, we present MRAE, a hybrid architecture with a DenseNet121 convolutional encoder and two transformer encoders (ViT-Base and DeiT-Base-Distilled). These branches are run in parallel with input sizes of 192×192 pixels, 224×224 pixels, and 256×256 pixels, respectively. The learned feature representations from these branches are fused using a cross-attention block and combined using learnable ensemble weights. Focal loss with deep supervision is used during training along with CutMix regularization, Weighted Random Sampling, and Cosine Annealing. We perform experiments using 10-fold stratified cross validation on the benchmark BUSI dataset (780 Images). MRAE achieves an average accuracy, macro F1-score, and macro recall of 93.72%, 94.25%, and 95.02%, respectively, across all cross-validation folds. The ResNet50 baseline achieves accuracy, F1-score, and recall of 90.64%, 91.36%, and 91.62% across all folds. We show that MRAE has significantly lower standard deviations across cross folds, indicating better stability. Our method provides evidence that breast ultrasound images can be classified accurately and reliably in a multi-resolution attention fusion network for use in clinical breast cancer screening.

@article{Alshateri2026,
  title     = {MRAE: Multi-Resolution Attention Ensemble with Hybrid CNN–Transformer Fusion for Breast Ultrasound Classification},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {5},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Hemin Kareem Azeez Alshateri and Ahmed Harbaoui},
  doi       = {10.14569/IJACSA.2026.0170536},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170536}
}