Automated Image Guidance for Diagnosing Skin Cancer With Confocal Microscopy

 描述(申请人提供)：在美国，黑色素瘤被诊断出约有124,000人，每年导致约10,000人死亡。皮肤科医生依靠肉眼和皮肤镜检查来区分良性黑素细胞病变和恶性黑素细胞病变，导致从8：1到47：1的高和高度可变的良恶性活检比率，以及数百万不必要的良性病变活检。反射共焦显微镜(RCM)成像已经在几个大型临床研究中被证明是无创性地指导黑色素瘤的诊断。真皮-表皮交界处RCM成像的敏感性为92-88%，特异性为71-84%。其特异性是皮肤镜检查的2倍。DEJ的RCM成像现在正在实施，以排除恶性肿瘤，减少活检和指导治疗。然而，这目前只在少数几个地点进行，那里有训练有素的专家，他们可以确保适当地进行成像和正确读取图像。这些专家是一小部分“早期采用者”临床医生，他们在过去十年中从事过RCM技术的工作，并已成为高技能的读者。对于更广泛的队列中热衷于采用RCM的新手(非专家)临床医生来说，学习阅读图像是具有挑战性的，需要大量的努力和时间。两个主要的技术障碍是新手临床医生在诊断准确率上的巨大差异的基础。它们共同限制了RCM的实用性、重复性和更广泛的采用。第一个是用户在获取图像的DEJ附近深度的主观可变性，第二个是图像解释的可变性。我们建议通过计算的“多方面”分类建模(创新)、图像分析和机器学习算法来解决这些障碍。我们的具体目标是：(1)开发和评估用于皮肤镜图像和RCM深度堆栈的算法，以实现在DEJ获得黑素细胞病变的RCM马赛克的自动化标准化和一致性；(2)开发和评估算法，将DEJ的细胞形态模式区分为两类，良性病变和恶性病变(发育不良病变和黑色素瘤)；以及(3)在患者身上测试我们的算法以获取RCM马赛克，并将其分类为这两组，与病理进行统计验证，并与病理进行统计验证。初步研究表明，我们的 算法可以准确地描绘出深色皮肤的DEJ在~3-13μm和浅色皮肤的DJ在5-20μm的范围内，并可以检测细胞形态模式，灵敏度在67-80%，特异度在78-99%。在DEJ上，黑色素细胞性皮损可以与周围的正常皮肤区分开来，分类准确率为80%。我们是来自纪念斯隆-凯特琳癌症中心、东北大学和摩德纳大学的一组研究人员。我们的成功将产生标准化的成像和分析方法，推动RCM在黑色素瘤的非侵入性检测中的应用。此外，这些方法对非黑色素瘤皮肤癌、皮肤淋巴瘤和其他皮肤病(影响更广)也很有用。

 DESCRIPTION (provided by applicant): Melanoma is diagnosed in approximately 124,000 people and is responsible for about 10,000 deaths every year, in the USA. Dermatologists rely on visual and dermatoscopic examination to discriminate benign melanocytic lesions from malignant, resulting in high and highly variable benign-to-malignant biopsy ratios from 8:1 to 47:1, and millions of unnecessary biopsies of benign lesions. Reflectance confocal microscopy (RCM) imaging has been proven for noninvasively guiding diagnosis of melanoma in several large clinical studies. RCM imaging at the dermal-epidermal junction (DEJ) provides sensitivity of 92-88% and specificity of 71-84%. The specificity is 2 times superior to that of dermatoscopy. RCM imaging at the DEJ is now being implemented to rule out malignancy, reduce biopsy and guide treatment. However, this is currently at only a few sites, where there are highly trained experts who can ensure that imaging is appropriately performed and images are read correctly. These experts are a small international cohort of "early adopter" clinicians, who have worked with RCM technology during the past decade and have become highly skilled readers. For novice (non-expert) clinicians in the wider cohort who are keen to adopt RCM, learning to read images is challenging and requires substantial effort and time. Two major technical barriers underlie the dramatic variability in diagnostic accuracy among novice clinicians. Together they limit utility, reproducibility and wider adoption of RCM. The first is user dependent subjective variability in depths near the DEJ at which images are acquired, and the second is variability in interpretation of images. We propose to address these barriers with computational "multi-faceted" classification modeling (innovation), image analysis and machine learning algorithms. Our specific aims are: (1) to develop and evaluate algorithms for both dermatoscopic images and RCM depth-stacks, to enable automated standardized and consistent acquisition of RCM mosaics at the DEJ in melanocytic lesions; (2) to develop and evaluate algorithms to discriminate patterns of cellular morphology at the DEJ into two classes, benign lesions versus malignant (dysplastic lesions and melanoma); and (3) to test our algorithms on patients for acquisition of RCM mosaics and classification into those two groups, with statistical validation against pathology, with statistical validation against pathology. Preliminary studies show that our algorithms can delineate the DEJ with accuracy in the range ~3-13 μm in strongly pigmented dark skin and ~5-20 μm in lightly pigmented fair skin, and can detect cellular morphologic patterns with sensitivity in the range 67-80% and specificity 78-99%. Melanocytic lesions can be distinguished from the surrounding normal skin at the DEJ with 80% classification accuracy. We are a team of researchers from Memorial Sloan-Kettering Cancer Center, Northeastern University and University of Modena. Our success will produce standardized imaging and analysis approaches, to advance RCM for noninvasive detection of melanoma. Furthermore, these approaches can be useful for non-melanoma skin cancers, cutaneous lymphoma and other skin disorders (wider impact).