An Adaptive Two-Stage Deep Learning Framework for Efficient and High-Sensitivity Breast Tumor Classification in Ultrasound Images

Abstract

Deep Learning models have demonstrated expert-level performance in classifying breast ultrasound images; however, state-of-the-art architectures often suffer from high computational complexity, rendering them unsuitable for deployment on resource-constrained portable medical devices. Furthermore, existing lightweight models typically prioritize inference speed at the expense of diagnostic sensitivity, a clinically unacceptable trade-off in cancer screening where false negatives can be fatal. To address these challenges, this paper proposes a Risk-Aware Adaptive Two-Stage Deep Learning Framework that dynamically balances computational efficiency with rigorous clinical safety. The framework utilizes a hierarchical architecture, employing EfficientNet-B0 as a rapid Stage-1 screener and DenseNet-121 as a robust Stage-2 specialist. Unlike standard adaptive networks that rely solely on entropy for routing, we introduce a novel Probability Risk Guard and an Aggressive Class-Weighted Training strategy. This ensures that any sample with even a marginal probability of malignancy is forwarded to the specialist model, preventing the premature dismissal of subtle tumor cases. Experimental validation on the Breast Ultrasound Images (BUSI) dataset demonstrates that the proposed framework achieves a Malignant Recall of 100%, successfully identifying all cancer cases in the test set, while maintaining an overall accuracy of 96%. Crucially, the adaptive gating mechanism successfully offloads 68.6% of input images to the lightweight Stage-1 model, significantly reducing average inference latency. These results confirm that the proposed framework offers a viable solution for real-time, high-sensitivity computer-aided diagnosis in clinical settings.