SCH: Leverage clinical knowledge to augment deep learning analysis of breast images

人工智能（AI）技术在基于医学图像的 应用.今天，在开发新的战略和方法方面存在前所未有的需求， 为各种应用程序提供强大、值得信赖和可访问的AI。经典的深度学习训练是 纯粹由数据驱动。在医学领域，临床知识往往是可用的和有用的，但大多数是 在当前的人工智能研究实践中被忽视了。将临床知识转化为深度学习建模 需要深入了解医疗背景/工作流程。这就需要多学科 利用计算技术和临床科学进行合作研究， 数据/AI研究。该项目的总体目标是开发一种新的深度学习范式， 结合成像数据和临床知识，以增强乳腺癌诊断，风险评估， 病变检测我们将在乳腺成像方面进行技术创新，以解决深度学习问题。 通过三个具体的方法，对小数据集、纵向检查和内容高效图像进行建模， 目的：目的1：将辅助任务/评估制定为用于乳腺癌诊断的CNN模型训练 目标2：利用图像的生物学关系指导深度学习结构设计 使用纵向数据进行乳腺癌风险预测;目标3：开发一个知识指导的无监督 用于识别可疑图的流水线，以支持对内容高效图像的深度学习分析。 这些目标代表了新的应用方法论发展，以建立粗糙的深度学习模型， 重要的临床成像应用。我们对每个目标都有很好的初步结果， 经验丰富的研究团队，涵盖计算，生物医学，工程和临床科学。我们 拟议的研究对开发强大和创新的人工智能战略/方法产生了更广泛的影响， 临床成像AI应用。超越乳房成像，我们提出的概念，范例， 这些方法也可以适用于其他疾病和成像模式，从而为治疗带来益处。 广泛的生物医学成像分析。获得的任何算法、知识、见解和经验 这一研究成果将对生物医学的快速发展和应用产生直接而实质性的影响。 医疗成像AI设备，最终使研究人员、临床医生和患者受益。

Artificial intelligence (AI) technologies have achieved remarkable success in medical image-based applications. Today, there are unprecedented needs in developing novel strategies and methodologies to enable robust, trustworthy, and accessible AI for various applications. Classic deep learning training is driven purely by data. In the medical domain, clinical knowledge is often available and useful, but is mostly ignored in the current practice of AI research. Incorporating clinical knowledge into deep learning modeling requires an in-depth understanding of medical context/workflow. This calls for multi-disciplinary collaborative research using computational techniques and clinical sciences to advance the biomedical data/AI research. The overall goal of this project is to develop a new paradigm of deep learning that combines imaging data and clinical knowledge to augment breast cancer diagnosis, risk assessment, and lesion detection. We will develop technical innovations on breast imaging to address deep learning modeling on small datasets, longitudinal examinations, and content-efficient images, through three specific aims: Aim 1: Formulate auxiliary tasks/assessment into model training of CNNs for breast cancer diagnosis on small datasets; Aim 2: Employ biological relationships of images to guide deep learning structure design for breast cancer risk prediction using longitudinal data; Aim 3: Develop a knowledge-guided unsupervised pipeline for identification of a suspicion map to support deep learning analysis on content-efficient images. These aims represent novel applied methodological development to build roust deep learning models for important clinical imaging applications. We have strong preliminary results for each aim and an experienced research team covering computational, biomedical, engineering, and clinical sciences. Our proposed study has a broader impact on developing robust and innovative AI strategies/methods to enable clinical imaging AI applications. Going beyond breast imaging, our proposed concepts, paradigms, and methods can also be adapted/applicable to other diseases and imaging modalities, leading to benefits for a wide range of biomedical imaging analyses. Any algorithms, knowledge, insights, and experience gained from this study will have a direct and substantial impact on the rapid evolvement and applications of medical imaging AI devices, ultimately benefiting the researchers, clinicians, and patients.