Statistical analysis and modelling of bi-modal autofluorescence-Raman imaging for efficient diagnosis and treatment of biological tissues

拉曼光谱是测量分子振动的一种特殊类型的光谱技术，已经成功地应用于在足够精细的尺度上确定生物组织的化学成分。这反过来又使得获得的信息可以很容易地用于皮肤的诊断和手术治疗，以及潜在的其他类型的癌症，如乳腺癌。该方法有望比现有的诊断和手术实践提供更准确、成本更低的替代方案，因此可以更快速、更广泛地获得这些类型的医疗保健。就患者体验而言，该方法还通过显著减少不必要的切除或其他创伤组织的数量，以及手术连续阶段之间的滞后，提供了改进。在其naïve实施中，该方法首先扫描整个生物样品，然后处理来自每个位点的光谱数据，以便随后进行自动分析，以确定样品中是否存在癌性形成，或者更雄心勃勃地在每个探测位置产生生物描述。然而，这种naïve实现需要非常长的时间，无法实现在一次不间断的手术中应用该方法的目的。提出了一种双模式成像解决方案，其中几乎瞬时的初步自体荧光成像结合自动聚类技术随后引导拉曼光谱仪将其测量集中在被认为更可能含有癌症的片段上。然后，在大量先前分析的样本上训练的统计模型试图完成每个部分的癌症检测任务，如果它对当前的诊断还没有足够的信心，就要求进行更多的拉曼测量。实施这种方法的设备现已在一家医院进行了试验，并准备在其他NHS中心以及国际上进行试验。虽然目前报告的采用该方法产生的结果令人鼓舞，但仍有几个方向可以推进该方法，并随后将现有技术转化为满足医疗保健提供者和患者期望的尖端最终产品。特别是，虽然医院的客观评价在很大程度上证实了该方法的预期灵敏度为90%，但特异性（非癌样本被正确诊断为非癌的比例）明显低于预期。此外，虽然目前的技术需要长达30分钟才能做出诊断，但最终目标是5-10分钟。我们提出了先进的统计和计算模型和方法，利用以前未充分利用的空间和形态信息来实现所需的转换。提出的方法还包括将目前使用的通用多元统计分析升级为功能数据分析，该分析利用拉曼光谱的内在功能结构，从而从分析的生物组织中提取更准确的生化标记。非欧几里得统计方法也被认为可以有效地捕获和表示光谱数据中的空间变化，并随后使用这些空间信息更准确地识别组织类型。最近可用数据的扩大也将使我们能够利用更复杂的统计分类模型来捕捉组织类型之间的细微差异，从而导致更准确和更可靠的癌症检测。我们来自爱沙尼亚（欧盟）的研究伙伴也支持这项建议，他们将通过调查另一类统计模型来补充我们的工作，增加交付高价值最终产品的总体机会。

Raman spectroscopy, a particular type of spectroscopic technique to measure molecular vibrations, has been successfully applied to determine chemical composition of biological tissues at sufficiently fine scales. This has in turn allowed the acquired information to be readily used for diagnosis and surgical treatment of skin, and potentially other types of cancer, such as breast cancer. The approach promises a more accurate and significantly less costly alternative to the existing diagnosis and surgical practices, and therefore a more rapid and broader access to these types of healthcare. In terms of patient experience the approach also offers an improvement through a significant reduction of both, the amount of unnecessarily removed or otherwise traumatised tissue, and the lag between successive stages of the procedure. In its naïve implementation, the method would first scan the entire biological sample before processing the spectral data from each site for subsequent automated analysis to establish presence or absence of cancerous formations in the sample or, more ambitiously, to produce a biological description at each probed location. However, this naïve implementation takes prohibitively long time, defeating the purpose of applying the method during a single uninterrupted surgery. A bi-modal imaging solution has been proposed, in which a nearly instantaneous preliminary autofluorescence imaging combined with an automated clustering technique subsequently guides the Raman spectrometer to concentrate its measurements on segments deemed more likely to contain cancer. A statistical model trained on a large number of previously analysed samples then attempts to complete the cancer detection task in each segment, requesting more Raman measurements if it is not already sufficiently confident in the current diagnosis. A device implementing this methodology has now been trialed in one, and is ready to be trialed in other NHS centres, as well as internationally. While the currently reported results produced by the current implementation of the methodology are encouraging, there still remain several directions for advancing the methodology and subsequently transforming the existing technology to a cutting edge final product that will meet the expectations of the healthcare providers and patients. In particular, while the objective evaluation in the hospital largely confirms the expected sensitivity of the method as 90%, the specificity (proportion of non-cancer samples correctly diagnosed as non-cancer) is notably lower than expected. Also, while currently the technology requires upto 30 minutes to produce a diagnosis, the ultimate aim is 5-10 minutes. We propose advanced statistical and computational models and methods, which utilize previously under-utilized spatial and morphological information to achieve the required transformation. The proposed methods also include upgrading currently used generic Multivariate Statistical Analysis to Functional Data Analysis, which takes advantage of the intrinsic functional structure of the Raman spectra and hence extracts more accurate biochemical markers from the analysed biological tissues. Methods of non-Euclidean statistics are also considered to efficiently capture and represent spatial variation in the spectral data and subsequently use such spatial information for more accurate recognition of the tissue types. The recent enlargement of the available data shall also allow us to take advantage of more complex statistical classification models capturing finer differences between the tissue types and subsequently leading to more accurate and robust detection of cancer. The proposal is also supported by our research partners from Estonia (EU) who are going to complement our work by investigating an additional class of statistical models, increasing the overall chance of delivering a highly valuable final product.