Histological examination is needed to diagnose pharyngeal and laryngeal carcinoma and intraoperative aid assessment of free margins during oncologic surgery. Various optical methods, such as confocal laser endomicroscopy (CLE), optical coherence tomography, and near-infrared fluorescence endoscopy, have been shown to perform an optical biopsy, i.e., differentiate benign from malign lesions.1–5 CLE provides a magnifying power up to 1000 times by using fluorescein to outline intercellular spaces and depict tissue architecture and has been extensively applied in the evaluation of mucosa lesions in gastroenterology, urology, and pneumology.6–11 CLE image interpretation has a moderate learning curve, and regarding the interrater reliability, there is inconsistency in the literature.11–14 In order to gain acceptance in the clinical setting, however, result consistency is needed. A classification system was developed and validated in the head and neck region to the oral mucosa, but there is no data regarding this scoring sheet in the pharyngeal or laryngeal region.15 This scoring system prioritizes homogeneity and intercellular gap changes in the tissue architecture and the inconsistency of cell morphology. Fluorescein leakage and vessel shape are also taken into consideration.15 Our main objective was to evaluate its diagnostic metrics in identifying pharyngeal and laryngeal carcinoma compared to the gold standard of histopathological examination.

We report on the transferability of a CLE classification system15 developed for oral mucosa to the pharyngeal and laryngeal epithelium. Based on 197 sequences (11.820 images) blinded to the gold standard of histopathological examination, differentiation of benign and malign lesions was possible with an accuracy, sensitivity, and specificity, of 83.2%, 81.3%, 85.5%, respectively. The examination added around 15 min to standard operation time without any side effects after intravenous fluorescein application. Tissue architecture (2 points), intercellular gaps (2 points), cell morphology (2 points), fluorescence leakage (1 point), and vessel shape (1 point) were classified according to the DOC-Score. In this regard, we confirmed the 2.6 mm CLE probe provided sufficient spatial resolution and tissue contrast to distinguish cellular architecture, borders, and size to characterize healthy and malignant tissue in the intraoperative setting. Malignant tissue samples were classified on average with 6.2 points and benign epithelium with 2.0 points among all examiners. Interpretation of CLE images are subjective and a crucial factor in making a correct assessment that requires, up to now, the experience and expertise of specialists. In this study, we defined experts as having performed more than 20 examinations with CLE. Interrater variability has been described with varying results (fair to almost perfect), especially when non-experts performed the examination.11–14

Oetter et al. developed and validated the DOC-Score based on 95 sequences (6224 images) to diagnose oral HNSCC through CLE.15 Using the DOC-Score, sensitivity and specificity of 95.3% and 88.9% for experts and 97.3% and 88.1% for non-experts was achieved. Inter-rater reliability (Fleiss` kappa) was 0.73 for experts and 0.814 for non-experts. Our values were lower (Sensitivity and Specificity of 83.2%, 82.1%, respectively) with an AUC of 0.88 for experts and 0.82 for non-experts. Interrater reliability was 0.73 for non-experts in our group, which was above the 0.64 calculated for the whole group and thus demonstrates the value of a systematic classification in a CLE-naive examiner. An accurate histopathological diagnosis of HNSCC is vital in guiding clinical management, and gold-standard for intraoperative assessment of free margins are adequate circumferential frozen sections. CLE images have a similar resolution to traditional H&E and could serve as valuable adjunct technology by potentially minimizing the need for intraoperative frozen section analysis. Pathological assessment of frozen sections and biopsy specimens does not typically involve the surgeon; however, a trained head and neck surgeon can perform CLE image interpretation for a real-time distinction of normal from abnormal tissue. The ability to sample the entirety of a tumor and surrounding tissue at the time of surgery may help guide the procedure's extent of resection. A pathologist (P), blinded to the histopathological examinations, also evaluated the CLE images. Interestingly the pathologist achieved an accuracy of 85.8%, very much comparable to the expert group (S1, S2, S3: 84.8%–88.8%) and better than the non-expert (S4, S5: 76.1%–77.7%), which underlines how principles used to classify histopathological slides are in some form adaptable to CLE images. Due to the physical properties of the technique, the penetration depth is fixed by 65 µm. Therefore, it was impossible to distinguish carcinoma in situ from invasive carcinoma since stromal invasion cannot be used as a criterion to differentiate between these two.21

A further possibility to objectivize CLE findings is utilizing automatic classification.22–24 Aubreville et al. showed that an approach based on transfer learning from intermediate endpoints within a pre-trained inception v3 network with preprocessing could reach an overall 94.8% accuracy, significantly improving overall performance over the traditional state of the art feature-based machine learning approaches.22 Automatic classification methods for CLE in the head and neck were developed for vocal cords and oral mucosa, but until this point, there is no available data for the pharyngeal and remaining laryngeal region. A prerequisite in developing such an automatic classification method based on deep learning-based approaches is acquiring large amounts of data and the correct labeling of such images in the healthy epithelium of a specific anatomic region and cancerous cells. Interestingly, compared to a previous study on the transferability of automatic classification based on algorithms trained with images of vocal cords applied on the oral cavity and vice versa, an accuracy of 68.5% and 89.5% was found.24 This suggests that the epithelium of these areas and the carcinomas that arise are similar on CLE; however, there seem to be significant differences that limit its detection rate. The same difficulties can also be pointed to the blinded examiners, mostly in the oral cavity and on vocal cords/larynx. The present work increases the information and knowledge of, until now, less examined anatomic regions and confirms the clinical criteria used to classify the oral cavity lesions are also mostly valid in the pharyngeal and laryngeal regions. Some differences are assumed since the diagnostic metrics are admittedly good but still not as accurate as for the oral cavity.15 Starting from the DOC-Score basis, some adaption of these criteria is probably required if it is to be applied to the whole head and neck region and should be assessed in further studies.

The absence of classification criteria and small data available limit CLE’s clinical application in head and neck until this day. However, we demonstrated that the DOC-Score provides a very good basis for establishing the head and neck region's diagnosis, despite the better performance in the oral cavity, which was the region it was originally developed for. CLE has, however, as a method with experimental status, limitations intrinsic to this technology and independent of the classification system used. As the probe-based CLE has only a penetration depth of 55–65 μm, it is not possible to assess the deeper margins in muscle and fascia. Indeed, mucosal epithelium is at its thinnest (floor of the mouth) over 100 μm, which formally makes the differentiation between carcinoma in situ and invasive carcinoma, as well as infiltration of submucosa, muscle and fascia very difficult.11,25 Direct evaluation of these deeper layers with CLE will need to be investigated in the future. CLE has a considerable potential to aid intraoperative characterization of mucosal regions in the head and neck, dependent on further validation in clinical trials.

By differentiating in real-time, in vivo healthy mucosal cells from squamous cell carcinoma with a sensitivity and specificity of 81% and 86%, respectively, CLE is a promising imaging technology that may improve the non-invasive characterization of HNSCC. A standardized classification system could improve diagnostic accuracy and consistency of results among examiners.

This project was supported by the German Research Foundation (DFG, Deutsche Forschungsgemeinschaft) with grant number GO 3182/2-1, MA 4898/17-1, OE 743/1-1, STE 1877/7-1; Project Number 439264659.

The authors declare no conflicts of interest.