Multimodal confocal microscopy for surgical guidance of skin resections

摘要 莫氏显微外科手术是治疗非黑色素瘤皮肤癌最有效的方法。莫斯 通过术中冰冻切片评估手术切缘获得高成功率(98%的治愈率) 组织病理学。不幸的是，执行冻结操作所需的复杂基础设施和繁琐流程 切片组织病理学导致漫长而昂贵的手术，限制了获得机会，并导致护理差距。 我们建议开发“光学Mohs”作为一种快速、低基础设施的替代方案，用于Mohs指征的患者 目前没有接受Mohs手术的农村和其他服务不足的人口。我们的光学Mohs 将基于多模式共焦显微镜(MCM)与机器学习相结合的方法来提供 低基础设施、自动化诊断工具，只需最少的组织处理。MCM联合收割机 将反射、荧光和拉曼共焦显微镜集成到单个台式平台中。MCM(使用 反射和荧光)最近展示了成功地产生未经处理的、 新鲜切除的皮肤，病理学家可以与冰冻切片组织病理学相媲美的准确性。 然而，这种方法本身仍然需要病理学家来阅读图像。人们正在探索机器学习 自动诊断这些图像，但尚未产生足够的准确性。我们假设 拉曼光谱的加入将显著提高自动化的诊断准确率 接近。拉曼是一种补充方法，它对皮肤的分子组成很敏感，并具有 已在临床边缘检测研究中得到证实，敏感性为92%-100%，特异性为 84-93%；然而，采用它的一个关键障碍是它的获取速度很慢。我们介绍两个 拉曼采集方面的创新(超像素和行扫描)，可实现高速拉曼采集 与手术指导兼容(速度为1cm2/min)。我们的初步模型在30名患者中显示 根据结构反射共聚焦图像和生化信息训练的预测模型 从拉曼图像提取区分基底细胞癌和正常结构 精确度，这表明光学Mohs可以帮助皮肤科医生根据需要“不断切割”，以移除 在不切除过多健康组织(92%特异性)的情况下，整个肿瘤(100%敏感性)。我们 将设计、制造和台架测试MCM仪器(目标1)。我们将设计一个决策支持系统，用于 使用108名患者的术后数据集，基于MCM图像进行肿瘤边缘评估(目标2)。我们 将决定基于MCM的肿瘤切缘评估决策支持系统的准确性 72例术中环境下的影像检查(目标3)。潜在的临床结果将证明 光学Mohs引导手术可用于常规Mohs手术，但目前不适用 使用，将Mohs的精确度扩大到目前没有接受他水平的护理的人群。

Abstract Mohs micrographic surgery (Mohs) is the most effective method to treat nonmelanoma skin cancer. Mohs achieves high success (98% cure rates) by assessing surgical margins intraoperatively with frozen section histopathology. Unfortunately, the sophisticated infrastructure and laborious process needed to perform frozen section histopathology leads to lengthy, expensive surgeries that limit access and result in disparities of care. We propose to develop "optical Mohs" as a rapid, low-infrastructure alternative for Mohs-indicated patients in rural and other underserved populations who do not currently undergo Mohs surgery. Our optical Mohs approach will be based on multimodal confocal microscopy (MCM) combined with machine learning to provide a low infrastructure, automated diagnostic tool requiring minimal tissue processing. MCM combines reflectance, fluorescence, and Raman confocal microscopy into a single benchtop platform. MCM (using reflectance and fluorescence) has recently demonstrated success in producing H&E images of unprocessed, freshly excised skin that pathologist can read with accuracy comparable to frozen section histopathology. However, this approach alone still requires a pathologist to read the image. Machine learning is being explored to automate the diagnosis of these images, but has not yet yielded sufficient accuracy. We hypothesize that the addition of Raman spectroscopy will significantly increase the diagnostic accuracy of an automated approach. Raman is a complementary approach that is sensitive to the skin’s molecular composition and has been proven in clinical margin detection studies within the skin with sensitivities of 92-100% and specificities of 84-93%; however, a critical barrier to its adoption has been its slow acquisition speed. We introduce two innovations in Raman acquisition (superpixel and line scanning) that enable acquisition of Raman at speeds compatible with surgical guidance (speeds of 1cm2/min.). Our preliminary model in thirty patients demonstrates that a predictive model trained on both structural reflectance confocal images and biochemical information extracted from Raman images discriminates basal cell carcinoma from normal structures with very high accuracy, suggesting that optical Mohs could help dermatologists "keep cutting" as needed to remove the entire tumor (100% sensitivity) while not removing an excessive amount of healthy tissue (92% specificity). We will design, fabricate and bench-test an MCM instrument (Aim 1). We will design a decision-support system for tumor margin assessment based on MCM images using a post-surgery data set in 108 patients (Aim 2). We will determine the accuracy of the decision support system for tumor margin assessment based on MCM imaging in an intraoperative setting in 72 patients (Aim 3). The potential clinical outcome would demonstrate that an optical Mohs guided surgery could be used where conventional Mohs is indicated but not currently used, expanding access of Mohs’ accuracy to populations currently not receiving his level of care.