目的 本研究旨在构建一个基于时序的超声深度学习模型，以期在乳腺癌新辅助治疗（neoadjuvant therapy，NAT）前无创预测乳腺癌分子分型，明确NAT早期超声评估疗效的最佳时机。 方法 纳入广州市第一人民医院接受完整NAT的乳腺癌患者176例，根据模型的两个目标将患者分成两组：176例（用于分型预测）和167例（用于疗效预测），收集其病理学结果及系列超声图像，根据免疫组化将肿瘤分为四种分子分型；根据术后病理的Miller-Payne分级将疗效分为显著缓解及无显著缓解。采用U-Net-Efficient-Net-B0混合架构建模，并引入“分割引导注意力”机制（SegAttend-Net），通过分析治疗前及NAT早期不同时期肿瘤随的声像图变化，深度提取特征，从而实现对分子分型及NAT疗效的预测。采用Clopper-Pearson精确法和Bootstrap法计算模型预测指标的置信区间，用Benjamini-Hochberg法进行多重对比校正模型在各治疗周期中对疗效预测的各性能指标，通过混淆矩阵及性能指标时序变化图评估模型效能。 结果 在分子分型预测任务中，模型对Luminal A型、Luminal B型、HER-2过表达型和三阴型预测准确率为82%、88%、72%和96%。在疗效预测任务中，预测的准确率从第1治疗周期的71%提升至第4治疗周期的80%，敏感度从0.14提升至0.79，并且在第3 — 4治疗周期的预测中其敏感度的提升具有统计学意义。 结论 该研究构建的U-Net-EfficientNet-B0混合模型能够在NAT前有效预测乳腺癌分子分型，并在NAT早期对疗效评估展现出临床实用价值，其中以第4治疗周期对疗效预测的效果最佳。

Objective This study aimed to develop a longitudinal deep learning ultrasound model to achieve two objectives： non-invasive prediction of breast cancer molecular subtypes prior to neoadjuvant therapy（NAT） and identification of the optimal timing for early efficacy assessment during NAT. Methods We enrolled 176 breast cancer patients from Guangzhou First People's Hospital who completed the full NAT course. The cohort was stratified into two analysis subsets： 176 patients for molecular subtyping and 167 for treatment response evaluation. Pathological data and serial ultrasound images were collected. Tumors were categorized into four molecular subtypes via immunohistochemistry. Treatment response was classified as "significant" or "non-significant" based on postoperative Miller-Payne grading. We employed a hybrid U-Net-EfficientNet-B0 architecture integrated with a segmentation-guided attention mechanism （SegAttend-Net）. The model leveraged pre-treatment images and dynamic sonographic feature changes across early NAT stages to predict molecular subtypes and therapeutic response. Confidence intervals were calculated using Clopper-Pearson exact and Bootstrap methods. Performance metrics across treatment cycles were adjusted for multiple comparisons using the Benjamini–Hochberg procedure. Evaluation utilized confusion matrices and longitudinal performance trajectories. Results In molecular subtyping， the model achieved accuracies of 82% （Luminal A）， 88% （Luminal B）， 72% （HER2-overexpressing）， and 96% （triple-negative）. For efficacy prediction， overall accuracy increased from 71% at cycle 1 to 80% at cycle 4， while sensitivity improved markedly from 0.14 to 0.79. The sensitivity improvement between the 3rd and 4th cycles was statistically significant. Conclusions The developed SegAttend-Net model demonstrates efficacy in pre-NAT molecular subtyping and holds clinical value for early efficacy assessment， with optimal predictive performance observed at the fourth treatment cycle.