Addressing algorithmic and data challenges to deep learning based segmentation of spine anatomy

项目摘要/摘要 在脊柱医学中，对生物医学图像的主观解释经常导致错误的诊断，旷日持久 对于可以手术治疗的患者进行非手术治疗，如果不需要手术治疗，则进行手术治疗。客观化 使用深度学习的上述图像的计算机化分析具有改进外科手术的潜力 结果，同时通过取消不必要的手术和加快必要的手术来降低手术成本 一个。然而，有几个障碍阻碍了深度学习技术的开发和部署，使之无法运作 外科实践中基于影像生物标记物的治疗建议。首先，一个公开可用的数据库是 缺席帮助培训和验证脊柱病理学的算法。第二，深度学习技术仍然存在 由于各种挑战，难以在临床环境中进行培训和操作。这些问题包括：1.不足 在基于深度学习的分割中将泛化误差与训练数据联系起来的框架，由于 算法的性能评估在部署之前是站不住脚的2.缺乏一种有纪律的方法来 提高深度网络在医学图像分割上的性能；3.缺乏支持 深入的网络，以确定和标记应咨询人类专家的疑难案件和失败案件。 首先，我们建议发展一个可供公众查阅的脊柱影像资料库，以推动深部医学的发展。 学习算法。第二，我们的目标是通过以下方式解决上述技术挑战：1.开发 基于幂定律标度的训练样本量与泛化误差解析关联框架 以及验证创建从深度学习模型到深度学习模型的深度学习集成的数学框架 保证改进分割性能3.开发和验证冯-诺伊曼信息- 基于分数，赋予深度学习集合识别疑难病例和预测失败的能力。

Project Summary/Abstract In spine medicine, subjective interpretation of biomedical images often leads to wrong diagnoses, prolonged non-surgical treatment for surgically treatable patients, and surgical treatment when none is necessary. Objective computerized analysis of the aforementioned images using deep learning has the potential to improve surgical outcomes while driving down the cost of surgery by eliminating unnecessary surgery and expediting necessary ones. Yet several barriers stymie the development and deployment of deep learning technology to operationalize imaging biomarker-based treatment recommendation in surgical practice. First, a publicly available database is absent to help train and validate algorithms for spinal pathologies. Second deep learning techniques remain difficult to train and operationalize in the clinical setting, due to various challenges. These include – 1. The lack of a framework to link generalization error to training data in deep learning-based segmentation, due to which performance estimates of algorithms are untenable prior to deployment 2. the lack of a disciplined approach to improve deep network performance on medical image segmentation and 3. the lack of frameworks that enable deep networks to identify and flag a difficult case and failed cases where a human expert should be consulted. First, we propose to develop a publicly accessible spine imaging database to promote the development of deep learning algorithms. Second, we aim to address the aforementioned technical challenges by 1. Developing a power-law scaling based framework to link training sample size and generalization error analytically 2. Proposing and validating a mathematical framework to create deep learning ensembles from deep learning models to guarantee improvement in segmentation performance 3. Developing and validating a Von-Neumann information- based score to endow deep learning ensembles with the ability to identify difficult cases and predict failure.