基于多模态超声自监督对抗学习分析的乳腺微小癌诊断及风险预测

乳腺占位（≤1cm）病变检出率逐年升高，但确诊率低且临床尚无有效的无创预测患者预后理论基础。课题组前期实验证明多模态超声结合深度学习对乳腺微小癌的诊断准确性高，有望解决这些临床问题。目前拟解决的关键科学问题是：如何通过多模态超声自监督对抗学习分析提高乳腺微小癌的确诊率；多模态超声是否能进行无创预测患者预后，并提供理论基础。不同于传统思路主要利用大量样本和单一模态影像进行肿瘤诊断，本项目拟通过挖掘乳腺微小癌的超声影像学特点，将其诊断过程与超声多模态影像自监督对抗学习融合，探究如何对小样本数据集提取有效表达特征；研究如何从大量直径1cm以上乳腺占位的临床图像中学习的模型知识高效迁移到微小癌模型中的关键技术；解决如何基于深度学习理论进行超声多模态影像分析以提高乳腺微小癌确诊率和是否可以预测转移、复发等风险科学问题，并探讨相关机制研究。为提高确诊率，减少过度活检数量及改善预后提供新理论依据、方法。

Despite the progress achieved in detection breast lesions (≤1cm), the confirmed diagnosis rate remains low, and currently there is no suitable theoretical guidance for non-invasive prediction of patient prognosis. Our previous work suggests that multimodal ultrasound combined with deep learning has a high accuracy in the diagnosis of breast microcarcinoma, which is expected to solve these clinical problems. The key scientific problems are: the multi-scale training model based on breast tumors has a low diagnostic rate in microcarcinoma，the application of deep learning in new ultrasonic technologies such as multi-mode and multi-scale is insufficient now，and the small number of existing cases leads to difficulty in deep learning. This project aims to explore how to extract effective representative features from small sample data sets by exploring ultrasonic imaging features of breast microcarcinoma and integrating its diagnosis process with ultrasonic multimodal image self-supervised admissive learning. Secondly, how to efficiently transfer the model knowledge learned from a large number of clinical images of breast mass with a diameter of more than 1cm to the microcarcinoma model. Finally, to solve the scientific problems of how to improve the diagnosis rate of breast microcarcinoma and whether it can predict the risk of metastasis and recurrence by using ultrasonic multimodal image analysis based on deep learning theory. A new theoretical basis is also proposed for increasing the diagnostic rate, reducing the number of excessive biopsy and improving the prognosis.