Skip to content
Conference Open access

Quantization-Aware Training of MobileNetV2 for Efficient Breast Ultrasound Image Classification

Sep 2026 · 2026 International Conference on Data Science and Its Applications (ICoDSA) · pp. 84-89 · 0 citations · 34 references

Abstract

Breast cancer is one of the leading causes of cancer death in women worldwide, making accurate and efficient early detection systems an urgent clinical need. Breast ultrasound (US) images offer a safe and affordable modality compared to mammography, but are susceptible to speckle noise and inter-observer variability that hamper consistent clinical interpretation. Although various deep learning approaches have shown promising diagnostic performance, most studies rely on computationally inefficient high-parameter architectures, while studies specifically integrating quantization techniques for breast US image classification in multi-dataset scenarios remain very limited. This study implements MobileNetV2 optimized via Quantization-Aware Training (QAT) for three-class breast US image classification using a combined BUSI and Mendeley dataset comprising 1,028 images after data cleaning. The model was trained in full-precision FP32 as a baseline, followed by QAT fine-tuning and INT8 conversion. The baseline FP32 model achieved an F1-Score of 92.71%, an accuracy of 93.55%, a precision of 91.61%, and a recall of 94.01%. After INT8 conversion, the model successfully maintained an F1-Score of 92.01% with only a 0.70% decrease, while achieving a 3.50× compression ratio, equivalent to a 71.4% reduction in model size, and $\mathbf{1. 7 3} \times$ CPU inference speedup. These findings confirm that QAT is an effective optimization strategy that successfully balances diagnostic performance and computational efficiency for heterogeneous dataset-based breast US image classification.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.