A holistic comparative study of deep learning architectures for automated breast ultrasound screening

Breast cancer screening through ultrasound imaging plays a crucial role in early detection, but manual interpretation is time-consuming and there are inter-observer differences. This study implemented and evaluated 12 cutting-edge deep learning architectures on the publicly available BUSI dataset (including benign, malignant and normal ultrasound images). This includes visual Transformer variants (such as MViTv2, ViT-EVA, BEiT-V2) and modern convolutional networks (such as EfficientNetV2-RW, ConvNeXt). To address the issue of category imbalance, we designed a targeted data augmentation process for categories with insufficient samples and adopted an early stop training strategy based on the selection of the validation set model. We reported the comprehensive metrics (accuracy, F1 score), model complexity (number of parameters, number of floating-point operations), avg time per epoch, and visualized the trade-off relationship between computational cost and classification performance. The results show that although some Transformer-based models have a relatively high inference cost, they can achieve an accuracy rate comparable to or better than that of the convolutional baseline. The lightweight convolutional architecture, on the other hand, demonstrates better efficiency. These findings provide practical guidance for the deployment of automated breast ultrasound screening in resource-limited clinical Settings.