Curating musculoskeletal CT data to enable the development of AI/ML approaches for analysis of clinical CT in patients with metastatic spinal disease

项目摘要/摘要 脊椎骨转移，广泛存在于癌症患者中，破坏脊椎解剖、骨结构、 和机械强度，使患者面临病理性脊椎骨折(PVF)的高风险。PVFS原因 高达50%的患者出现神经功能障碍，进一步的并发症可能是致命的。我们的父母资助 (AR075964)开发并验证了一种新的计算肌肉骨骼方法， 临床CT对PVF风险的准确预测。脊椎解剖、骨骼属性和 来自临床CT成像的单个脊柱肌肉横截面积，组装的基本步骤 由于癌症介导的改变，计算性肌肉骨骼模型面临着独特的挑战 骨骼组织的放射学表现。深度学习(DL)方法的应用将加快和标准化 关键的分段步骤，允许分析更大的数据集，从而促进新的DL分析以实现改进 洞察脊柱转移性疾病患者PVF风险的驱动因素。这一发展受到了 缺乏公开可用的、有临床注释的特定于转移性人类脊椎的图像数据。 这项提议旨在建立一个可供公众访问的、经过管理的人类转移瘤的4D CT成像数据集 脊柱和相关的病变和脊椎特征的放射学描述，以促进 分析PVF风险的DL方法。因此，我们提出了三个具体目标：1)为 深度学习的应用：我们的数据是基于我们的父母赠款(AR075964)去识别的脊柱和 胸腹肌CT图像数据集从140名患者在放疗计划和其他 随访3个月、6个月、9个月和12个月的CT诊断数据(共1400个图像数据集)。我们将集结并 管理用于公开分发的数据，方法是a)为机器学习应用程序对CT图像进行预处理，b) 描绘图像中的脊椎、病变和肌肉，c)汇编元数据，包括有限的主题 将人口统计和疾病状况输入JSON文件，以及d)提取SIFT特征以进行计算机视觉风格的分析。 2)测试床深度学习(DL)分割：确保精选的数据集适合人工训练 智能和机器学习(AI/ML)系统，我们将开发和训练试验床DL分段网络 在基线和随访的临床CT中对骨骼、病变和肌肉进行分割。我们将使用网络来控制 策划的CT图像和勾画的质量，3)按照最佳做法传播4D数据集： 完成任务1和任务2后，我们将通过癌症图像向研究社区提供数据 归档(TCIA)遵循其已建立的用于取消识别DICOM扫描以及注释和编码的方法 临床资料及分析结果。在我们的方法中集成数字图书馆系统将改变患者 管理模式从被动管理到数据驱动的主动式管理，以预防PVF事件 减少病人管理中的偏见。

Project Summary/Abstract Vertebral bone metastases, widespread in patients with cancer, destroy vertebral anatomy, bone architecture, and mechanical strength, exposing patients to a high risk of pathologic vertebral fracture (PVF). PVFs cause neurological deficits in up to 50% of patients with further complications that may be fatal. Our parent grant (AR075964) develops and validates a novel computational musculoskeletal approach for patient-specific, precise prediction of PVF risk from clinical CT. Segmentation of vertebral anatomy, bone properties, and individual spinal musculature cross-sectional area from clinical CT imaging, a fundamental step for assembling the computational musculoskeletal models, faces unique challenges due to the cancer-mediated alteration in skeletal tissues radiological appearance. Application of deep learning (DL) methods will speed and standardize the critical segmentation step, permitting analysis of larger datasets promoting new DL analysis for improved insight into the drivers of PVF risk in patients with metastatic spine disease. This development is hindered by the lack of publicly available, clinically annotated image data specific to metastatic human spines. This proposal aims to establish a curated, publicly accessible, 4D CT imaging dataset of human metastatic spines and associated radiological delineations of lesions and vertebral features, to enable the advancement of DL methods to analyze PVF risk. We thus propose three specific aims: 1) prepare the CT dataset for the application of deep learning: Our data is based on our parent grant (AR075964) de-identified spinal column and thoracoabdominal muscles CT image datasets obtained from 140 patients at radiation planning and additional CT diagnostic data at 3, 6, 9, and 12 months follow up (1400 image data sets overall). We will assemble and curate the data for public distribution by a) preprocessing the CT images for machine learning applications, b) delineating the vertebrae, lesions, and muscles in the images, c) assembling metadata, including limited subject demographic and disease status into JSON files, and d) extract SIFT features for computer-vision style analyses. 2) Testbed Deep Learning (DL) Segmentation: To ensure that the curated data set is suitable for training artificial intelligence and machine learning (AI/ML) systems, we will develop and train testbed DL segmentation networks to segment bones, lesions, and muscles in baseline and follow-up clinical CT. We will use the networks to control the quality the curated CT images and delineations, 3) Disseminate 4D dataset following best practices: Upon completion of tasks 1 and 2, we will make the data available to the research community via the Cancer Image Archive (TCIA) following their established methods for de-identifying DICOM scans and annotating and encoding clinical data and analysis results. Integrating DL systems within our approach will change the patient management paradigm from reactive to data-driven proactive management to prevent PVF events and critically reduce bias in patient management.