A combined spectral-spatial framework for the diagnosis and stratification of colorectal cancer through advanced fast Raman imaging

背景：伦敦大学学院细胞与发育生物学和病理学部门与雷尼绍PLC之间的合作越来越多，以开发新的疾病诊断方法，首先是结肠直肠癌（CRC）。这种疾病是英国卫生系统的一个主要负担。2008年，英国在诊断结直肠癌阳性病例上花费了2000多万英镑。这支撑了2亿英镑用于初级手术、化疗和放射治疗。随后的复发管理使这些疾病的费用又增加了4亿英镑。此外，NHS花费了超过2.5亿英镑用于诊断首次发现没有结直肠癌的患者。已经确定，与其他欧盟国家相比，英国的生存率较低，关键是这与疾病在诊断和获得治疗时的阶段直接相关。组织病理学：目前诊断结直肠癌的金标准，仍然主要是一种主观技术。这在确定早期疾病患者时导致了重大问题，这些患者可以更保守地治疗，而且成本效益更高。但不幸的是，病理学家对早期肿瘤的看法非常不一致。因此，更准确客观的方法来识别那些将迅速发展为晚期疾病的患者是至关重要的。拉曼分析以可复制的方式从未标记的组织中提供疾病特异性分子信号。这开启了在疾病早期肿瘤阶段进行临床决策的可能性，目前的金标准显示只有50%的观察者之间的一致性（在发育不良分类中，保守治疗可能是最有效的）。该项目的及时性和重要性一位通过UCL Impact计划资助的博士生已经建立了拉曼成像在人类结肠活检组织的拉曼化学成像图中区分早期病理特征的能力。然而，该项目的瓶颈现在是在数据分析和建立能够支持拉曼显微镜图像自动诊断的引人注目的统计模型的水平上。这就是一个具有强大跨学科前景的初级研究数学家、统计学家或计算科学家的技能可以产生深远影响的地方。诊断模型必须考虑到在组织样本中逐像素获得的拉曼光谱以及结肠癌形态变化的几何特征。每一种方法本身都是一个挑战，但我们的目标是将光谱中的高维信息分析与光谱空间分析中特征化学特征的几何分布结合起来。其目标是提供一种硬件和软件组合来监测结直肠癌的疾病进展，并最终达到MHRA和美国FDA等监管机构的标准。项目的技术方面：在拉曼显微地图中呈现的高维和高度结构化的信息带来了一些有趣的数据分析挑战。它既可以看作是一组随机场，也可以看作是一个多变量过程的大集合。相关性存在于空间和光谱域的不同尺度上，除了非平稳性/异质性之外，还提供了许多独特而有趣的方法机会。除了光谱成分的低维表示（通过PCA或其他方式），该项目还将探索所谓的“特征袋”方法的效用，以得出数据中存在的空间相互作用的分布。预计这种双重空间光谱框架应该比传统的自动病理检测方法具有显著的优势。

Background:There is a growing collaboration between the Departments of Cell & Developmental Biology and Pathlology at UCL and Renishaw PLC to develop new diagnostic approaches to disease and in the first instance Colorectal cancer (CRC). This disease is a major burden on the UK health system. In 2008 the UK spent over £20M on diagnosis of positive cases of CRC. This underpinned >£200M spent on primary surgical, chemo and radiotherapies. Subsequent management of recurrence added a further £400M to the costs of these diseases. Additionally, the NHS spent over £250M in diagnosing patients found to be free of CRC in the first pass. It has been established that the UK survival rate is poor compared to other EU countries and crucially that this correlates directly with stage of the disease at diagnosis and access to treatment. Histopathology: the current gold standard for CRC diagnosis, is still mainly a subjective technique. This leads to significant issues when identifying patients in early stage disease who could be treated more conservatively and very much more cost effectively. But unfortunately agreement between pathologists for early neoplasias can be very poor. Consequently, more accurate objective methods of identifying those patients who will progress rapidly to advanced disease is vital. Raman analysis provides disease specific molecular signals from unlabelled tissues in a reproducible manner. This opens up the possibility of clinical decision making at early neoplastic stages of disease, where currently the gold standard is shown to have only 50% inter-observer agreement (in classification of dysplasias, where conservative treatment is likely to be most effective).The timeliness and importance of this projectA PhD student funded through the UCL Impact scheme has established the power or Raman imaging to distinguish early pathological features in Raman chemical imaging maps of biopsy tissue from the human colon.However the bottleneck in the project is now at the level of data analysis and the building of compelling statistical models capable of supporting automated diagnosis from the Raman microscope image. This is where focusing the skills of a junior research mathematician, statistician or computational scientist with a strong interdisciplinary outlook can have a profound impact. The diagnostic models must take account of the Raman spectra obtained pixel-by-pixel across the tissue sample as well as the geometric features characteristic of the morphological changes of colon cancer. Each of these would be a challenge in their own right but our ambition is to combine the analysis of the very high dimensional information from the spectra with the geometric distribution of characteristic chemical signatures in a combined spectral-spatial analysis. The aim is to deliver a hardware and software combination to monitor CRC disease progression that can ultimately meet the standards of regulators like MHRA and the US FDA. The technical aspects of the project: The high-dimensional, and highly structured information, present in Raman microscopic maps carry some fascinating data analysis challenges. It can be regarded as both a set of random fields and as a large collection of multivariate processes. Correlations, which exist at different scales across both the spatial and spectral domain, in addition to the non-stationarity/heterogeneity, present many unique and interesting methodological opportunities. In addition to low-dimensional representations of the spectral components (via PCA or otherwise), this project will also explore the utility of so-termed "bags-of-features" approaches to elicit a distribution of spatial interactions present in the data. It is anticipated that this dual spatial-spectral framework should hold significant advantages over traditional approaches to automated pathology detection.