Applied artificial intelligence for medical imaging

动力。人工智能(AI)是一种很有前途的促进医疗实践的工具。采用广泛使用的方法，如卷积神经网络，在临床支持中发挥重要作用是具有挑战性的，在许多情况下，它将是特定的应用程序。我的研究项目旨在设计、开发和部署人工智能技术，以帮助医务人员检测和诊断不同的神经疾病。方法：研究方法。我提出了三个相互关联的项目，这三个项目将推动人工智能在神经成像领域的应用。在项目1中，我们将改进基于深度学习的神经退行性疾病分类算法。目前，标准的方法是使用在大型自然图像数据集上训练的现有深度学习体系结构，然后根据医学数据进行微调。由于自然图像和神经学数据之间的差异，这种方法并不能得到最优的结果。我们将使用直接基于神经学数据训练的更简单的卷积结构。该项目的结果将是一个MRI特定的神经网络，然后将通过转移学习适用于其他神经成像数据集。项目2，这个项目涉及设计新的计算机视觉算法来识别脑肿瘤。我们将专注于胶质瘤，这是一种通常在MRI图像中弥漫出现的肿瘤，这使得它们的分割成为一项具有挑战性的任务。该项目的目标是创建能够识别胶质瘤的软件，并提供有关其位置、大小和类型的信息，以及在纵向数据中量化肿瘤的变化。该系统将帮助医疗专业人员识别肿瘤的具体变化，避免目前主观和容易出错的肉眼检查。项目3将通过创建从新生儿到成年的大脑生长轨迹模型来推进新生儿MRI数据纵向分析的算法。我将通过创建从1岁到80岁的每一年的大脑模板来扩展我之前在新生儿成像方面的工作，以创建MRI发育工具箱。该项目将为临床研究提供一个软件工具箱，其中将包括多年龄模板、注册和建模功能。这个目前尚不存在的工具箱将有助于识别退化轨迹，为药物设计和治疗评估提供相关信息。冲击力。这项提案中的研究将提高人工智能在医学成像中的适用性，成为计算机科学和医疗系统之间的桥梁。作为Enigma-Aaxia集团的联席主管，我将从这个项目(全球21个地点)分发我们的软件，供医院测试和采用。我们的软件将减少专业人员的工作量，使神经学诊断更快、更准确。因此，提高患者将从卫生保健系统获得的治疗质量。

Motivation. Artificial intelligence (AI) is promising tool to advance medical practice. Adapting widely used approaches such as convolutional neural networks to play a significant role in clinical support is challenging and, in many cases, it will be application specific. My research program seeks to design, develop, and deploy AI techniques to assist medical personnel in the detection and diagnosis of different neurological conditions. Methods. I propose three interconnected projects that will advance the application of artificial intelligence in the field of neuroimaging. In Project 1, we will improve algorithms for neurodegenerative diseases classification using deep learning. Currently, the standard approach is to use an existing deep learning architecture trained on a large natural image dataset, and then fine--tune it to the medical data. Due to the differences between natural images and neurological data, this approach does not lead to optimal results. We will use simpler convolutional architectures trained directly on the neurological data. The outcome of this project will be a MRI-specific neural network that will then be applicable to other neuroimaging data sets via transfer learning. Project 2, This project involves the design of novel computer vision algorithms to identify brain tumors. We will focus on gliomas, a type of tumor that usually appear diffused in MRI images making their segmentation a challenging task. The goal of this project is to create software that can identify gliomas and provide information regarding their location, size, and type, as well as quantification of the tumor change in longitudinal data. The system will aid the medical professional to identify specific changes in the tumor, avoiding the current subjective and error-prone visual inspection. Project 3 will advance the algorithms for longitudinal analysis of neonatal MRI data by creating a brain-growth trajectory model from neonate to adulthood. I will extend my previous work in neonatal imaging by creating brain templates from each year of life from 1 to 80 years of age to create the MRI developmental toolbox. This project will provide a software toolbox for clinical research that will include multi-age templates, registration, and modeling functions. This currently non-existent toolbox will help to identify degeneration trajectories providing relevant information for drug design and treatment evaluation. Impact. The research in this proposal will improve the applicability of AI in medical imaging, acting as a bridge between computer science and the health system. As co-director of the ENIGMA-Ataxia group, I will distribute our software to be tested and adopted in the hospitals from this project (21 sites across the world). Our software will reduce the workload of specialized professionals making neurological diagnoses faster and more accurate. Hence, improving the quality of treatment that patients will receive from the health-care system.