Improved Diagnosis of Shunt Malfunction with Automatic Quantification of Ventricular Space

摘要 脑积水是脑脊髓液（CSF）在脑深部的空腔（脑室）中积聚。的 脑积水的最常见治疗是通过脑室腹膜（VP）分流术进行CSF分流。超过 据估计，美国每年放置30，000个VP分流管。尽管这种情况 手术进行时，并发症发生率估计为近24%，有一份报告引用了22% 修订率。近50%因分流相关问题入院的患者需要住院5天或以上。 考虑到手术部位感染率和与分流探查和翻修相关的并发症， 分流器故障的准确诊断对于许多神经外科医生来说仍然是一个关键的（如果难以实现的话）目标。之一 建立仅基于成像的诊断的困难在于缺乏标准化的稳健方法， 测量心室大小。最近，体积分析已被研究作为测量的方法， 心室大小，与Evans指数或额枕角比率相比， 更准确和更好的工具来测量心室大小对分流的反应。但 在分割心室时相关的人的努力以及观察者之间和观察者内部的可变性阻碍了其广泛的应用 临床采用。诊断脑室扩大或脑积水的另一个困难是， 缺乏一个标准化的、规范的数据集，其中包括各种被认为是“正常”的数据集。 年龄范围是心室大小随年龄增加而增加。目前的文献缺乏一个强大的规范性数据集， 按年龄和性别划分的心室大小，直到最近才为儿科年龄产生这样的数据集 范围在所有年龄人群中建立心室容积和形态的标准值 迫切需要，并将允许调查各种主题相关的脑积水，并最终 协助检测和分类脑积水和VP分流相关并发症或故障。在 近年来，深度学习（DL）模型的快速发展对许多领域产生了巨大的影响， 医学，特别是自动图像分析任务，包括分割。利用DL 本研究的主要目的是：1）建立一个具有鲁棒性的心室DL模型 包括多模态支持和自动故障检测的分割，并建立规范的数据库; 2)开发一个软件原型，该原型包含DL模型和规范值，并适合临床 分流器故障的基于图像的诊断工作流程。最终，一个独特的软件产品将 开发并商业化，以改善分流故障和脑积水的诊断， 手术效果好，费用低。

ABSTRACT Hydrocephalus is the buildup of cerebrospinal fluid (CSF) in the cavities (ventricles) deep within the brain. The most common treatment for hydrocephalus is CSF diversion via ventriculoperitoneal (VP) shunting. Over 30,000 VP shunts are placed per year in the United States by some estimates. Despite how commonly this surgery is performed, the complication rate has been estimated at almost 24%, with one report citing a 22% rate of revision. Nearly 50% of patients admitted with shunt related issues require a stay of five or more days. Given the rate of surgical site infections and complications associated with shunt explorations and revisions, accurate diagnosis of a shunt malfunction remains a critical, if elusive, goal for many neurosurgeons. One of the difficulties in establishing a diagnosis based on imaging alone is the lack of standardized robust methods of measuring ventricular size. Recently volumetric analyses have been studied as a method for measuring ventricular size, as compared to the Evans’ Index or frontal-occipital horn ratios and have been suggested is more accurate and a better tool for measuring response of ventricular size to shunting. However, the associated human efforts and inter- and intra-observer variability in segmenting the ventricles prohibits its wide clinical adoption. The other difficulty with establishing a diagnosis of ventriculomegaly or hydrocephalus, involves a lack of a standardized, normative dataset with a range of what is considered "normal" for various age ranges as the ventricle size increases with age. Current literature lacks a robust normative dataset of ventricular size by age and gender and only recently has such a dataset been produced for the pediatric age range. Establishment of normative values for ventricular volume and morphology across all age population is sorely needed and will allow for the investigation of a variety of topics related to hydrocephalus and ultimately assisting in the detection and triage of hydrocephalus and VP shunt related complications or malfunctions. In recent years, the rapid development of deep learning (DL) models has led to great impact on many areas of medicine, especially for automatic image analysis tasks including segmentation. Taking advantage of DL models, two aims are proposed in this project: 1) develop and validate a robust DL model for ventricle segmentation including multi-modality support and automatic failure detection and build a normative database; 2) develop a software prototype that incorporates the DL model and normative values and fits the clinical workflow for image-based diagnosis of shunt malfunction. Ultimately, a unique software product will be developed and commercialized to improve the diagnosis of shunt malfunction and hydrocephalus and benefit the patients with better surgical outcome and reduced cost.