恶性肿瘤患病率和致死率在逐年递增，乳腺癌的患癌率和致死率在女性中均排世界第一。对乳腺癌常见的两种治疗方式是手术治疗和新辅助治疗。其中，评估术前淋巴结转移与否和转移负荷的方法与评估新辅助化疗是否达到达到病理完全缓解（pCR）的方法都是有创检查，增加了患者心理、身体负担和治疗成本。计算机辅助诊断模型已经在各类单模态、单任务的场景下，为临床提供了良好的支撑。本研究针对乳腺癌手术和新辅助治疗两种方式，以高通量影像数据、多角度病理结果和高维度基因检测等多模态数据为输入，探索多模态融合、缺失模态补全、小样本学习和多任务学习等问题中的隐含信息表征、临床信息插值、最优化建模、模型参数迭代更新等数学理论和工程方法，分别构建乳腺癌淋巴结转移负荷和手术后生存期多任务预测系统和乳腺癌新辅助治疗pCR状态和治疗后生存期多任务预测系统。本研究方法将具有鲁棒性、可解释性和可泛化性，有助于辅助乳腺癌在临床上的诊疗。

The diagnostic rate and mortality rate of malignant tumors are increasing in recent years. Breast cancer is the most typical cancer for women around the world. Breast surgery and neoadjvant therapy (NAT) are two main treatment methods for breast cancer. In clinic, both the method to evaluate lymph node metastasis and metastasis burden before surgery, and the method to evaluate pathological Complete Response (pCR) after NAT are invasive, which increase the psychological, physical and financial burden of patients. Currently, single-task and single-modality computer-aided diagnosis models have been developed and played an important role in clinic. Based on this technology, this study will focus on the scenarios of breast surgery and NAT, analyze multi-modality clinical data including high-throughput imaging data, multi-phase pathological data, and high-dimension genetic data all together, explore the mathematical methods of hidden feature representation, clinical information interpolation, optimization modeling and parameter iteration for sub-tasks of modality fusion, missed modality generation, few-shot learning and multi-task learning. A joint-prediction system for lymph node metastasis, metastasis burden and survival of patients undergoing surgery and a joint-prediction system for pCR and survival of patients undergoing NAT will be established. These systems will be robust, interpretable and generalizable, which potentially can assist the doctor in clinic.