随着信息技术的发展，医疗数据的采集和存储水平不断提高，深度学习技术在医学领域有着广阔的应用前景。深度学习技术可以从电子医疗数据中提取有效信息，辅助医生提高诊断效率与精度。如何对不同类别电子医疗数据进行有效的分析利用，并探索实际临床应用场景成为技术应用的挑战。针对现有的技术问题，本文对深度学习技术在静态和时序影像数据、三维影像数据及多模态数据等不同数据层面的疾病辅助诊断问题展开研究。首先，本文从器官表型角度，利用超声影像进行肝脏肿瘤诊断分析研究。针对超声影像的异质性和现有静态超声图像诊断模型的局限性，本文提出了基于超声影像分割信息的肿瘤引导策略，以及基于注意力提升机制的双向卷积长短期记忆诊断模型，通过提取并融合不同超声视频帧之间的特征信息，充分学习肝脏肿瘤区域的形态信息及其与周围组织的关系，提高模型诊断准确性。本文从影像和临床指标角度解释模型决策，并探索用双读机制辅助医生提高诊断精度。其次，本文从面部表型角度，利用三维面部影像进行年龄、生活习惯和代谢疾病预测分析研究。本文利用深度卷积神经网络提取面部全局特征，通过定位关键点、划分面部感兴趣区域，提取面部局部纹理和形态特征。之后，本文将面部全局特征与局部特征进行融合，开发联合模型，分析面部特征与年龄、生活习惯与代谢疾病的相关性。另外，本文还探索了模型在智能手机上的应用前景。最后，本文从多表型角度，利用多模态电子医疗数据进行女童中枢性性早熟疾病诊断分析。本文利用电子病历、实验室检查、超声与射线数据进行女童中枢性性早熟疾病预测。针对该场景下的模态缺失问题，本文提出了动态多模态变分自编码器模型，学习不同模态特征间的关联性，补齐缺失模态特征。实验结果表明，利用由动态多模态变分自编码器补齐的缺失模态特征，分类器能取得更精确的预测结果。同时，本文也对不同模态特征在诊断过程中的作用进行了样本层面以及全局层面的分析解释，探索了模型在临床中的实用价值。综上所述，本文从多种临床问题研究场景出发，研究针对不同类型的电子医疗数据的深度学习方法，为解决临床医学应用问题提供助益。

With the development of information technology and the increasing level of medical data collection and storage, deep learning technology has a broad application prospect in the medical field. Deep learning technology can extract effective information from electronic medical data to assist doctors in improving diagnostic efficiency and accuracy.How to effectively analyze and utilize different medical data and explore application scenarios become challenges for clinical application. To address this issue, this study explored the disease diagnosis, and applies deep learning to static and time-series image data, three-dimensional image data in three clinical scenarios.First, from the perspective of organ phenotype, this study used ultrasound images to analyze the diagnosis and analysis of liver tumors.To address the heterogeneity of ultrasound images and the limitations of existing static-image based diagnostic models, this study proposed a mass-guide strategy based on ultrasound segmentation information and an attention-boosted bi-directional convolutional long short-term memory diagnostic model based on ultrasound videos, to extract and fuse features between different video frames, and learn morphological information of liver mass regions and their relationship with surrounding tissues, thus increasing diagnostic accuracy. This study explained the model decision from the perspective of images and clinical factors, and investigated to use the double-reading process to assist physicians to improve diagnostic accuracy.Secondly, from the perspective of facial phenotype, this paper used three-dimensional facial images to predict age, living habits and metabolic diseases. This study used deep convolutional neural networks to extract facial global features, and local texture and morphological features by locating landmarks and segmenting facial regions of interest (ROIs). A joint model was constructed for age prediction, lifestyle prediction, and metabolic disease classification by integrating global features and local features. A prospective pilot study was also conducted using 3D images taken from a smartphone to test the AI performance for clinical applications.Finally, from the perspective of multi-phenotype, this study used multi-modal electronic medical data to diagnose central precocious puberty in girls. Electronic medical records, laboratory tests, ultrasound and radiological data were used to predict central precocious puberty.To address the missing-modality problem, this study proposed a dynamic multimodal variational autoencoder to learn the correlation between different modalities, and impute features for missing modalities. The experimental results demonstrated that our models can impute features for missing modalities effectively and achieve higher diagnostic accuracy. This study also analyzed the feature attributions at the instance level and global level, and explored the practical value of the model in the appropriate clinical setting.In summary, this study investigated deep learning methods for different types of electronic medical data from various clinical scenarios, which can provide benefits for solving clinical problems.