Optical Biopsy for Tissue Diagnostics of Squamous Cell Carcinoma in the Upper Aerodigestive Tract using Confocal Laser Endomicroscopy Imaging

鳞状细胞癌(SCC)占口腔和咽部所有癌症类型的90%以上，以及几乎100%的喉部恶性肿瘤。目前，诊断的金标准是对组织进行侵入性活组织检查，然后进行组织病理学评估。激光共聚焦显微镜(CLE)是一种已成功应用于可疑粘膜病变目视检查的非侵入性方法。利用这种活体成像方法，激光被发射并施加到选定深度的组织上，然后将光的反射荧光重新聚焦以进行检测。荧光素通过静脉注射并通过细胞间隙分布，而不通过细胞膜扩散，从而使细胞组织的轮廓可视化和结构分析成为可能。由于其使细胞结构可见的特性，CLE据说可以提供实时的光学活组织检查。CLE检查高度依赖于检查人员的经验，在以前的出版物中显示了可变的诊断指标。这促使我们的团队引入了一种新的基于深度学习的自动分类方法。应用这种新方法，我们在交叉验证场景中达到了88.3%的准确率，此外，我们还证明了将从口腔上皮CLE图像中获得的分类知识转移到同一数据集中的声带上皮CLE图像上，准确率甚至提高了89.45%。这些结果表明，从这两个解剖位置获得的CLE成像数据可以帮助建立一个通用模型，可以区分这两个领域中的任何一个领域的正常组织和恶性肿瘤。作为对技术现状和我们之前工作的分析的结果，我们打算在这一申请资助期(24个月)内收集CLE图像数据，以提高数据质量和数量，这适合于医生的培训以及机器学习算法。该数据库将包括良性/恶性粘膜病变以及上呼吸道的生理性粘膜，并将以匿名数据为基础作为开放获取数据发布。有了这个丰富的数据集，我们打算通过使用深度学习方法，通过稳健地检测图像伪影并在临床水平上对恶性和良性结构变化执行细粒度分类，来进一步改进自动分类系统中的最先进技术。

Squamous cell carcinoma (SCC) accounts for over 90 percent of all cancer types in the oral cavity and pharynx, as well as for almost 100 percent of malignancies in the larynx. At present, the gold standard of diagnosis is an invasive biopsy of the tissue with subsequent histopathological assessment. One non-invasive method that has been successfully applied for visual inspection of suspicious mucosal lesions is Confocal Laser Endomicroscopy (CLE). With this in vivo imaging method, laser light is emitted and applied to tissue at a selected depth, the reflected fluorescence of light then being refocused for detection. Fluorescein is administered intravenous and distributed through the intercellular spaces without diffusing through the cell membranes, thus enabling outline visualization and structural analysis of cellular tissue. Due to its property of making cellular structures visible, CLE is said to provide ’real-time’ optical biopsies. CLE examination is highly dependent on examiners´experience and showed in previous publications variable diagnostic metrics. This motivated our group to introduce a new approach based on deep learning of automatic classification.Applying this new approach, we reached accuracies of 88.3% in a cross-validation scenario and , additionally, we demonstrated that it is possible to transfer classification knowledge acquired from epithelial CLE images of the oral cavity to epithelial CLE images of the vocal folds in the same data set, with even increased accuracies of 89.45%. These results suggest that CLE imaging data acquired from both of these anatomical locations can help to establish a general model that can distinguish normal tissue from malignancies in either of the two domains. As a result of the analysis of the state of the art and our own previous work, we intend to collect CLE image data during this applied funding period (24 months) in order to increase data quality and amount, which is suitable for the training of physicians as well as machine learning algorithms. This database will include benign/malignant mucosal lesions as well as physiological mucosa of the upper aerodigestive tract and will be released as open access data based on anonymized data. With this enriched data set, we intend to further improve on the state-of-the-art in automatic classification systems, by robustly detecting image artifacts and performing fine-grained classification of malign and benign structural changes at clinical level, both using deep learning approaches.