Automated Assessment for Robotic Suturing Utilizing Deep Learning Algorithms

摘要 医疗差错是美国第三大死因，每年造成200亿美元的损失。 手术并发症占死亡人数和费用的三分之一。外科手术表现 直接影响患者的预后。前列腺癌是男性最常见的癌症，接受治疗 通过可能导致阳萎的手术(机器人辅助前列腺癌根治术(RARP))， 大小便失禁甚至死亡。需要可靠的客观评估技术手段。 在这个项目中，我们将重点评估外科医生在RARP期间的缝合技能通过虚拟 现实(VR)模拟。缝合是许多类型的手术中的一种常见技能，可以被追踪 与绩效指标相关，并与RARP后的患者结果相关。 在这项建议中，我们寻求首先确定关键的子步缝合操作和 成功实现这些目标所需的技术技能(目标1a)。此外，我们打算发展 通过分析原始运动学数据的自动化技能评估渠道(目标1b)， 视频(目标2b)和运动学/视频(目标2c)，来自VR模拟性能 创新的机器学习策略和基于深度学习的计算机视觉。 建议工作的主要区别是确定细粒度的子步骤有多好 进行缝合时的操作。参与这项研究的外科医生不仅将提供 数据通过他们的VR模拟性能，但也将贡献来自 RARP以建立外科医生技能、患者因素和外科手术之间的关系 结果。统计建模将衡量外科医生技能和患者的不同影响 影响患者预后的因素(目标3)。我们假设机器的创新应用 学习算法可以准确评估外科医生的技术技能，并可以进一步预测 相关临床结果的可能性。 拟议的工作将在VR中实现可扩展和可操作的反馈，使外科医生能够 具有将活体手术中的手术风险降至最低的宝贵知识。

ABSTRACT Medical errors are the third leading cause of death in the US at a cost of $20 billion annually. Surgical complications account for a third of these deaths and cost. Surgical performance directly impacts patient outcomes. Prostate cancer, the most common cancer in men, is treated with surgery (robot-assisted radical prostatectomy (RARP)) that can lead to impotence, incontinence, and even death. Reliable means of objectively assessing technique are required. In this project we will focus on assessing surgeon suturing skills during RARP through virtual reality (VR) simulation. Suturing is a common skill in many types of surgeries, can be tracked with performance metrics, and has been correlated with patient outcomes after RARP. In this proposal we seek to first determine the critical sub-step maneuvers of suturing and the technical skills necessary to achieve them successfully (Aim 1a). Further, we intend to develop an automated skills assessment pipeline through the analysis of raw kinematic data (Aim 1b), video (Aim 2b), and both kinematic/video (Aim 2c), from VR simulation performance by innovative machine learning strategies and deep-learning-based computer vision. The primary differentiator of the proposed work is determining how well granular sub-step maneuvers in suturing are performed. Surgeons participating in this study will not only provide data through their VR simulation performance, but will also contribute real patient data from the RARP to establish the relationship between surgeon skill, patient factors, and surgical outcomes. Statistical modeling will measure the differential impact of surgeon skill and patient factors to patient outcomes (Aim 3). We hypothesize that innovative application of machine learning algorithms can accurately assess surgeon technical skills, and can further anticipate likelihood of relevant clinical outcomes. The proposed work will enable scalable and actionable feedback in VR, empowering surgeons with valuable knowledge to minimize surgical risk in live surgery.