多模态超声VisTran-Attention网络评估早期子宫颈癌保留生育功能手术可行性

早期子宫颈癌育龄患者保留生育功能是国家生育政策下的新问题；但保留生育功能手术术前对病理类型及淋巴结转移情况的评估依赖于有创性检查，存在准确率低和创伤大的问题。结合项目组前期多模态超声临床应用及超声联合深度学习的研究基础，提出建立无创精准评估模型的设想：超声图像微观纹理特征差异可映射与病理类型及淋巴结转移情况相关的肿瘤异质性，通过构建多模态超声深度学习网络对图像进行识别与分类，从而搭建评估模型。针对肿瘤特性，本项目拟引入搭载可变形注意力模块的Vision Transformer深度学习方法，构建早期子宫颈癌术前多模态超声图像VisTran-Attention分类网络，验证该网络与肿瘤病理类型及淋巴结转移生物特性的相关性，建立早期子宫颈癌保留生育功能手术的精准无创评估模型。本项目设计的多模态超声分类网络，有望成为早期子宫颈癌无创性评估的突破点。

Fertility preservation in early-stage cervical cancer patients of reproductive age has become a new problem under the national fertility policy. However, the evaluation of pathological type and lymph node metastasis depends on preoperative invasive examination, which has problems of low accuracy and unavoidable trauma. Combining the clinical application of multimodality ultrasound and our research foundation of artificial intelligence ultrasomics, we put forward the idea of setting up a noninvasive accurate evaluation model: the differences among micro ultrasonic image texture features can mapping tumor heterogeneity, which is associated with both pathological type and lymph node metastasis. Thus, we may set up an evaluation model by constructing a multimodality ultrasonic deep learning network for image recognition and classification. In view of tumor characteristics, we plan to introduce Vision Transformer deep learning method equipped with deformable attention module to construct VisTran-Attention classification network based on multimodality ultrasound images of early-stage cervical cancer, and verify the correlation between this network and biological factors related to tumor pathological types and lymph node metastasis. The establishment of this preoperative evaluation model is expected to be a breakthrough point for noninvasive evaluation of early cervical cancer.