Classification of basal cell carcinoma in ex vivo confocal microscopy images from freshly excised tissues using a deep learning algorithm

     Depósito de Investigación de la Universidad de Sevilla https://idus.us.es/ This is an Accepted Manuscript of an article published by Elsevier in Journal of Investigative Dermatology (JID) , Vol. 1291 - 1299.e2 , on May 2022, available at: https://doi.org/10.1016/j.jid.2021.09.029 Elsevier Copyright © 2021 The Authors. Published by Elsevier Inc. All rights reserved . En idUS Licencia Creative Commons CC BY-NC-ND 1 Article Type: Original article 1 Article Title: CLASSIFICATION OF BASAL CELL CARCINOMA IN EX VIVO 2 CONFOCAL MICROSCOPY IMAGES FROM FRESHLY EXCISED TISSUES USING 3 A DEEP LEARNING ALGORITHM 4 Short Title: Deep learning for confocal ex vivo images 5 6 Mercedes Sendín-Martín1 (ORCID ID# 0000-0003-0156-381X), Manuel Lara-Caro1 (ORCID 7 ID# 0000-0001-7424-2000), Ucalene Harris2 (ORCID ID# 0000-0002-7845-7460), Matthew 8 Moronta2 (ORCID ID# 0000-0002-7845-7460), Anthony Rossi2 (ORCID ID# 0000-0003-9 2295-1934), Erica Lee2 (ORCID ID# 0000-0001-9670-3242), Chih-Shan Jason Chen2 (ORCID 10 ID# orcid.org/0000-0002-0977-2707), Kishwer Nehal2 (ORCID ID# 0000-0001-5791-7184), 11 Julián Conejo-Mir Sánchez1,3 (ORCID ID# 0000-0001-9108-9538), José-Juan Pereyra-12 Rodríguez*1,4 (ORCID ID# 0000-0001-6843-5877), Manu Jain*2,4 (ORCID ID# 0000-0002-13 6035-0825). 14 1. Hospital Universitario Virgen del Rocío, Dermatology Department, Sevilla (Spain). 15 2. Memorial Sloan Kettering Cancer Center, Dermatology Service, Department of Medicine, 16 New York (USA). 17 3. School of Medicine, University of Seville (Spain). 18 4. Weill Cornell Medicine, Dermatology Service, Department of Medicine, New York (USA). 19 * Both authors contributed equally. 20 21 Corresponding author: Manu Jain 22 Memorial Sloan Kettering Cancer Centre (MSKCC) 23 530 E74th St NY, NY 10021 24 Email: [email protected] 25 2 Funding sources: This research is funded by a grant from the National Cancer Institute 26 /National Institutes of Health (P30-CA008748) made to the Memorial Sloan Kettering Cancer 27 Center. 28 Manuscript word count: 3500. 29 Figures: 4; Tables: 2; Supplementary figures: 2; Supplementary video: 1; Supplementary 30 tables: 1. 31 Keywords: ex vivo confocal microscopy, digitally stained purple and pink images, basal cell 32 carcinoma, convolutional neural network, deep learning 33 Abbreviations Used: BCC is Basal Cell carcinoma, MMS is Mohs Micrographic Surgery, AI 34 is Artificial Intelligence, EVCM is Ex Vivo Confocal Microscopy, ROC is Receiver Operator 35 Characteristic, AUC is Area Under the Curve, CNN is Convolutional Neural Network. 36 3 ABSTRACT 37 Ex vivo confocal microscopy (EVCM) generates digitally colored purple-pink images similar 38 to H&E, without time-consuming tissue processing. It can be used during Mohs surgery for 39 rapid detection of basal cell carcinoma (BCC); however, reading EVCM images requires 40 specialized training. An automated approach using a Deep Learning algorithm to BCC detection 41 in EVCM images can aid in diagnosis. 40 BCCs and 28 negative (“not-BCC”) samples were 42 collected at Memorial Sloan Kettering Cancer Center to create three training datasets: 1) EVCM 43 image dataset (663 images), 2) H&E image dataset (516 images), and 3) a combination of the 44 two datasets. 7 BCCs and 4 negative samples were collected to create a EVCM test dataset 45 (107 images). The model trained with the EVCM dataset achieved 92% diagnostic accuracy, 46 similar to the H&E model (93%). The area under ROC was 0.94, 0.95, and 0.94 for EVCM, 47 H&E, and combination trained models, respectively. We developed an algorithm for automatic 48 BCC detection in EVCM images (comparable accuracy to dermatologists). This approach could 49 be used to assist with BCC detection during Mohs surgery. Furthermore, we found that a model 50 trained with only H&E images (which are more available than EVCM images) can accurately 51 detect BCC in EVCM images. 52 53 54 55 56 57 58 59 60 61 4 INTRODUCTION 62 Basal cell carcinoma (BCC) is the most common skin cancer accounting for ~2 millions of 63 cases annually in the United States alone [Rogers et al. 2015]. Biopsy, followed by 64 histopathology, is the gold standard for diagnosis and subtyping of BCCs (as aggressive or non-65 aggressive) for appropriate management — surgical treatment for aggressive tumors versus 66 topical treatment (non-surgical) for less aggressive BCCs. For aggressive or recurrent tumors, 67 especially those located in cosmetically sensitive sites such as the face, Mohs micrographic 68 surgery (MMS) is the treatment of choice with high cure rates [Van Loo et al. 2014]. However, 69 MMS is a tedious and time-consuming procedure as it involves careful removal of skin cancer 70 layer-by-layer. Each excised layer undergoes frozen sectioning and microscopic evaluation, 71 requiring up to 20-45 minutes. Often times multiple layers are removed to achieve complete 72 tumor clearance; thus, the entire surgical procedure may last for several hours [Keena and Que 73 2016], increasing patient’s waiting time, complications, and cost of the procedure. Additionally, 74 frozen sectioning can cause tissue destruction and create artefacts that may hinder the final 75 diagnosis. 76 To expedite the surgical procedure, various ex vivo optical imaging devices have been 77 developed [Bennàssar et al. 2013; Dalimier and Salomon 2012; Gareau et al. 2009; Karen et al. 78 2009]. These devices can rapidly image freshly excised tissues at “near-histopathological” 79 resolution, obviating the need for destructive and time-consuming tissue processing. 80 Ex vivo confocal microscopy (EVCM) is an emerging imaging technique that can evaluate 81 freshly (un-processed) excised whole-tissue samples without the need for tissue processing 82 (frozen sectioning). As there is no tissue processing involved, EVCM can image tissues rapidly 83 (less than a minute for a tissue measuring up to 2 cm), reducing the time for the Mohs surgery 84 and enabling real-time imaging in the surgical suite [Keena and Que 2016]. Furthermore, 85 EVCM creates digitally colored purple and pink images (similar to H&E images) by converting 86 5 fluorescence signal originating from the nucleus into purple color and reflectance signal from 87 the cytoplasm into pink color. These digitally colored images can be read by Mohs surgeons 88 trained in pathology [Mu et al. 2016]. Although EVCM has demonstrated an overall high 89 sensitivity and specificity (~ 90%) for detection of BCC [Gareau et al. 2009; Karen et al. 2009] 90 during MMS [Bennàssar et al. 2013], it is only been utilized in very few academic centers. We 91 believe that the integration of an automatic algorithm for the detection of BCC in EVCM images 92 could immensely aid Mohs surgeons, increasing adoption of this technology. Moreover, EVCM 93 technology may be useful for surgical pathologists and dermatologists to obtain faster results 94 from standard excisions [Bennàssar A et al. 2012; Debarbieux et al. 2015] or even for biopsies 95 of inflammatory skin lesions [Bağcı et al. 2019; Bağcı et al. 2021; Bertoni et al. 2018]. This 96 approach can achieve the goal of a real “bedside” pathology, similar to the ongoing integration 97 of this technique for assessment of non-dermatology specimens in surgical pathology 98 [Panarello et al. 2020]. 99 Artificial intelligence (AI) is currently transforming healthcare [Hinton 2018]. A popular AI 100 technique for image classification is Convolutional Neural Networks (CNNs), a deep-learning 101 approach inspired from the human brain. In CNNs, inputs such as images go through several 102 layers of artificial “neurons” before an output is finally rendered such as the diagnoses of those 103 images. CNN algorithms are being used in radiology and pathology [Topol 2019] to classify 104 images as neoplastic or non-neoplastic and have shown a proficiency at par or even exceeding 105 human performance [Campanella et al. 2019]. Likewise, in dermatology, CNNs have reported 106 comparable performance to an expert dermatologist in skin cancer diagnosis using clinical 107 [Esteva et al. 2017; Fujisawa et al. 2019, Han et al. 2018] and dermoscopy images [Brinker et 108 al. 2019; Codella et al. 2016; Haenssle et al. 2018; Haenssle et al. 2020]. Recently, CNN has 109 also been applied successfully to reflectance confocal microscopy (RCM) images to classify 110 skin lesions [Kose et al. 2020, Wodzkinski et al. 2019; Wodzkinski et al. 2020, Campanella et 111 6 al. 2021]. However, to the best of our knowledge, CNNs have not been developed and tested to 112 diagnose skin cancers in digitally colored EVCM images. 113 The goal of this study was to develop and test the performance of CNN algorithms for detecting 114 BCC in digitally colored EVCM images obtained from freshly excised tissues from Mohs 115 surgery. For this study, 40 BCCs and 28 negative (“not-BCC”) skin tissue samples were 116 collected from 42 patients to create three different image datasets to train CNN models: 1) an 117 EVCM image dataset with 663 images, 2) an H&E image dataset with 516 images, and 3) a 118 combination of the two datasets (EVCM and H&E image datasets) with 1179 images. The 119 performance of these 3 trained models were evaluated and compared on a separate test set (not 120 used in training), which comprised of 97 EVCM images created using 7 BCCs and 4 negative 121 (“not-BCC”) skin tissue samples were from 11 patients. 122 123 RESULTS 124 Patient demographics and lesion characteristics: 125 A total of 53 patients were enrolled in the study. Mean age was 61 years ( 13, range 36-95 126 years); 64% (34/53) were males and 36% (19/53) females. Majority of the lesions, 63.3% 127 (50/79) were located on the head and neck. A total of 47 BCCs were imaged including 18 128 nodular BCCs (nBCCs), 11 superficial BCCs (sBCCs), 10 infiltrative BCCs (iBCCs), 8 mix-129 subtype BCCs and 32 skin samples did not present BCC. Patient demographics and lesion 130 characteristics are detailed in the Supplementary Table S1. 131 132 Model performance: 133 The main outcome measures were sensitivity, specificity, positive predictive value (PPV), and 134 negative predictive value (NPV) (Table 1 and Supplementary Figure S1). 135 7 The model trained with the EVCM dataset (EVCM model) achieved 92% diagnostic accuracy, 136 similar to the H&E model (93%). Compared to the H&E model, the EVCM model had a higher 137 sensitivity (96% vs. 93%) but lower specificity (89% vs. 92%). The combined model had the 138 lowest diagnostic accuracy (86%) with a high specificity (92%; similar to H&E model) but the 139 lowest sensitivity (78%). Area under the curve (AUC) of receiver operating characteristics 140 (ROC) for diagnostic dichotomous classification was calculated for each the three training 141 datasets (Figure 1). The AUC was 0.94, 0.95, and 0.94 for EVCM, H&E and combination 142 trained models, respectively. 143 144 Gradient maps: 145 Gradient maps created with Grad-CAM [Selvaraju RR, 2016] highlighted important regions in 146 red color in the images for predicting the presence/absence of tumor after all the three trainings. 147 Figure 2 shows gradient map examples of a true positive (TP), a true negative (TN), a false 148 positive (FP), and a false negative (FN) tissue sample. For the TP example, the algorithm 149 identified even small BCC nodules (asterisks) as important areas for BCC prediction, while not 150 taking the hair follicle (arrow) as an important region. Likewise, for the TN example, sebaceous 151 gland was correctly identified as an important region within the image for the negative 152 prediction for tumor. We had a very few images with false positive and false negative results. 153 For FP example, sebaceous glands (arrowhead) and eccrine ducts (star) were detected as 154 important for BCC detection. On the other hand, a FN example where BCC nodules (asterisks) 155 were not considered important and the algorithm prediction was no-BCC. 156 157 DISCUSSION 158 BCC is the most common skin cancer worldwide [Leiter et al. 2014]. Although, BCC has a low 159 metastatic potential it can be locally invasive causing extensive tissue damage and loss of 160 8 regional function [Nehal and Bichakjian 2018]. MMS is a specialized surgical procedure 161 capable of achieving complete clearance of BCC, while maximizing normal tissue preservation, 162 making it a preferred treatment for recurrent BCCs and BCCs located on cosmetically sensitive 163 and functionally challenging sites, such as the face [Jain et al. 2017]. However, MMS is not 164 only a time-consuming surgery (due to frozen section analysis) but it is also an expensive 165 procedure that requires an extensive laboratory set-up and specialized surgeons and technicians. 166 Ex vivo confocal microscopy (EVCM) is an emerging imaging technique that generates 167 digitally colored purple and pink images similar to H&E, without any time-consuming tissue 168 processing [Mu et al. 2016; Schüürmann et al. 2019]. It can be used during Mohs surgery 169 (MMS) for rapid detection of residual basal cell carcinoma (BCC); however, reading EVCM 170 images requires specialized training. An automated approach to BCC detection in EVCM 171 images can aid in diagnosis. 172 Currently, AI is being implemented extensively in the field of dermatology and pathology for 173 the automated diagnosis of skin cancers and non-neoplastic lesions (psoriasis, atopic dermatitis 174 and onychomycosis) [Han et al. 2018] in clinical and dermoscopy images using CNN, a deep-175 learning algorithm [Esteva et al. 2017, Schüürmann et al. 2019]. In our study we used CNN for 176 detection of BCC in EVCM images. CNN was first used in skin cancer detection by Nasr-177 Esfahani et al. (2016) for the diagnosis of melanoma. They trained the algorithm with a small 178 dataset of 170 clinical images from melanocytic lesions. Similar to our study, the authors used 179 augmentation methods such as random rotation, and resizing of the images to increase the 180 number of images (from 170 original images to 6120 images) in the training dataset, which 181 yielded a sensitivity of 81% and a specificity of 80% in the diagnosis of melanoma. 182 CNNs has been applied to non-invasive in vivo imaging technique such as reflectance confocal 183 microscopy (RCM). Wodzinski et al. (2019) reported an accuracy of 91% in the diagnosis of 184 BCC in RCM images. Recently, Campanella et al. (2021) developed a deep learning model to 185 15 sampling the result to the image size to plot the original image with the heatmap, highlighting 335 the important regions (red color) for the model prediction. A subset of gradient maps is shown 336 in Figure 2. 337 338 Statistical analysis: 339 All statistical analyses was performed in R (v4.03) [R Core Team, 2020]. The ability to 340 discriminate the classes (“BCC” or “not-BCC”) inferred by the model was used to generate 341 ROC curves with 95% intervals, using the package “pROC” (version 1.17.0.1) [Robin et al. 342 2011]. ROC curves and 95% confidence intervals for sensitivity, specificity, positive predictive 343 and negative predictive values, and accuracy measures were generated. 344 345 DATA AVAILABILITY 346 The deep learning model that was trained to generate the presented results and the code is 347 available at https://github.com/boxyware/confocal-ex-vivo. 348 349 CONFLICT OF INTERESTS 350 The authors declare none. 351 352 ACKNOWLEDGMENTS 353 We would like to thank Dr. Kivanc. Kose, PhD at Memorial Sloan Kettering Center for his 354 critical review of the manuscript. 355 356 357 AUTHOR CONTRIBUTIONS STATEMENT 358 Conceptualization: MSM, MJ. Data curation: MSM, MLC, MM, UH, MJ. Resources: AR, EL, 359 16 CJC, NK. Software: MLC. Formal Analysis: MLC. Results analysis: MSM, MJ, MLC, JJPR. 360 Writing – original draft: MSM. Writing – review & editing: MSM, MLC, JCS, JPR, MJ. 361 362 ORCID IDs 363 Mercedes Sendín-Martín1 (ORCID ID# 0000-0003-0156-381X), Manuel Lara-Caro1 (ORCID 364 ID# 0000-0001-7424-2000), Ucalene Harris2 (ORCID ID# 0000-0002-7845-7460), Matthew 365 Moronta2 (ORCID ID# 0000-0002-7845-7460), Anthony Rossi2 (ORCID ID# 0000-0003-366 2295-1934), Erica Lee2 (ORCID ID# 0000-0001-9670-3242), Chih-Shan Jason Chen2 (ORCID 367 ID# orcid.org/0000-0002-0977-2707), Kishwer Nehal2 (ORCID ID# 0000-0001-5791-7184), 368 Julián Conejo-Mir Sánchez1,3 (ORCID ID# 0000-0001-9108-9538), José-Juan Pereyra-369 Rodríguez*1,4 (ORCID ID# 0000-0001-6843-5877), Manu Jain*2,4 (ORCID ID# 0000-0002-370 6035-0825). 371 372 373 374 375 376 377 378 379 380 381 382 REFERENCES 383 17 Bağcı IS, Aoki R, Krammer S, Ruzicka T, Sárdy M, French LE, Hartmann D. Ex vivo confocal 384 laser scanning microscopy for bullous pemphigoid diagnostics: new era in direct 385 immunofluorescence? J Eur Acad Dermatol Venereol. 2019 Nov;33(11):2123-2130. doi: 386 10.1111/jdv.15767. 387 388 Bağcı IS, Aoki R, Vladimirova G, Ergün E, Ruzicka T, Sárdy M, et al. New-generation 389 diagnostics in inflammatory skin diseases: Immunofluorescence and histopathological 390 assessment using ex vivo confocal laser scanning microscopy in cutaneous lupus 391 erythematosus. Exp Dermatol. 2021 May;30(5):684-690. doi: 10.1111/exd.14265. 392 393 Bennàssar A, Vilalta A, Carrera C, Puig S, Malvehy J. Rapid diagnosis of two facial papules 394 using ex vivo fluorescence confocal microscopy: toward a rapid bedside pathology. Dermatol 395 Surg. 2012 Sep;38(9):1548-51. doi: 10.1111/j.1524-4725.2012.02467.x. 396 397 Bennàssar A, Carrera C, Puig S, Vilalta A, Malvehy J. Fast evaluation of 69 basal cell 398 carcinomas with ex vivo fluorescence confocal microscopy: criteria description, 399 histopathological correlation, and interobserver agreement. JAMA Dermatol. 2013 400 Jul;149(7):839-47. doi: 10.1001/jamadermatol.2013.459. 401 402 Bertoni L, Azzoni P, Reggiani C, Pisciotta A, Carnevale G, Chester J, et al. Ex vivo 403 fluorescence confocal microscopy for intraoperative, real-time diagnosis of cutaneous 404 inflammatory diseases: A preliminary study. Exp Dermatol 2018;27:1152–9. 405 https://doi.org/10.1111/exd.13754. 406 18 Brinker TJ, Hekler A, Hauschild A, Berking C, Schilling B, Enk AH, et al. Comparing artificial 407 intelligence algorithms to 157 German dermatologists: the melanoma classification benchmark. 408 Eur J Cancer 2019;111:30–7. https://doi.org/10.1016/j.ejca.2018.12.016. 409 410 Campanella G, Hanna MG, Geneslaw L, Miraflor A, Werneck Krauss Silva V, Busam KJ, et 411 al. Clinical-grade computational pathology using weakly supervised deep learning on whole 412 slide images. Nat Med 2019;25:1301–9. https://doi.org/10.1038/s41591-019-0508-1. 413 414 Campanella G, Navarrete-Dechent C, Liopyris K, Monnier J, Aleissa S, Minas B, et al. Deep 415 Learning for Basal Cell Carcinoma Detection for Reflectance Confocal Microscopy. J Invest 416 Dermatol. 2021 Jul 12:S0022-202X(21)01437-8. doi: 10.1016/j.jid.2021.06.015. Epub ahead 417 of print. PMID: 34265329. 418 419 Codella N, Nguyen Q-B, Pankanti S, Gutman D, Helba B, Halpern A, et al. Deep Learning 420 Ensembles for Melanoma Recognition in Dermoscopy Images. IBM J Res Dev 2016;61. 421 422 Dalimier E, Salomon D. Full-field optical coherence tomography: A new technology for 3D 423 high-resolution skin imaging. Dermatology 2012;224:84–92. 424 https://doi.org/10.1159/000337423. 425 426 Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist-level 427 classification of skin cancer with deep neural networks. Nature 2017;542:115–8. 428 https://doi.org/10.1038/nature21056. 429 430 19 Debarbieux S, Gaspar R, Depaepe L, Dalle S, Balme B, Thomas L. Intraoperative diagnosis of 431 nonpigmented nail tumours with ex vivo fluorescence confocal microscopy: 10 cases. Br J 432 Dermatol. 2015 Apr;172(4):1037-44. doi: 10.1111/bjd.13384. 433 434 Fujisawa Y, Otomo Y, Ogata Y, Nakamura Y, Fujita R, Ishitsuka Y, et al. Deep-learning-based, 435 computer-aided classifier developed with a small dataset of clinical images surpasses board-436 certified dermatologists in skin tumour diagnosis. Br J Dermatol 2019;180:373–81. 437 https://doi.org/10.1111/bjd.16924. 438 439 Gareau DS, Karen JK, Dusza SW, Tudisco M, Nehal KS, Rajadhyaksha M. Sensitivity and 440 specificity for detecting basal cell carcinomas in Mohs excisions with confocal fluorescence 441 mosaicing microscopy. J Biomed Opt 2009;14:034012. https://doi.org/10.1117/1.3130331. 442 Haenssle HA, Fink C, Schneiderbauer R, Toberer F, Buhl T, Blum A, et al. Man against 443 Machine: Diagnostic performance of a deep learning convolutional neural network for 444 dermoscopic melanoma recognition in comparison to 58 dermatologists. Ann Oncol 445 2018;29:1836–42. https://doi.org/10.1093/annonc/mdy166. 446 447 Haenssle HA, Fink C, Schneiderbauer R, Toberer F, Buhl T, Blum A, et al. Man against 448 Machine: Diagnostic performance of a deep learning convolutional neural network for 449 dermoscopic melanoma recognition in comparison to 58 dermatologists. Ann Oncol 450 2018;29:1836–42. https://doi.org/10.1093/annonc/mdy166. 451 452 Haenssle HA, Fink C, Toberer F, Winkler J, Stolz W, Deinlein T, et al. Man against machine 453 reloaded: performance of a market-approved convolutional neural network in classifying a 454 broad spectrum of skin lesions in comparison with 96 dermatologists working under less 455 20 artificial conditions. Ann Oncol 2020;31:137–43. 456 https://doi.org/10.1016/j.annonc.2019.10.013. 457 458 Han SS, Kim MS, Lim W, Park GH, Park I, Chang SE. Classification of the Clinical Images 459 for Benign and Malignant Cutaneous Tumors Using a Deep Learning Algorithm. J Invest 460 Dermatol 2018;138:1529–38. https://doi.org/10.1016/j.jid.2018.01.028. 461 462 Han SS, Park GH, Lim W, Kim MS, Na JI, Park I, et al. Deep neural networks show an 463 equivalent and often superior performance to dermatologists in onychomycosis diagnosis: 464 Automatic construction of onychomycosis datasets by region-based convolutional deep neural 465 network. PLoS One 2018;13. https://doi.org/10.1371/journal.pone.0191493. 466 467 He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Proc. IEEE 468 Comput. Soc. Conf. Comput. Vis. Pattern Recognit., vol. 2016- December, IEEE Computer 469 Society; 2016, p. 770–8. https://doi.org/10.1109/CVPR.2016.90. 470 471 Hinton G. Deep learning-a technology with the potential to transform health care. JAMA - J 472 Am Med Assoc 2018;320:1101–2. https://doi.org/10.1001/jama.2018.11100. 473 474 Jain M, Rajadhyaksha M, Nehal K. Implementation of fluorescence confocal mosaicking 475 microscopy by “early adopter” Mohs surgeons and dermatologists: recent progress. J Biomed 476 Opt 2017;22:24002. https://doi.org/10.1117/1.JBO.22.2.024002. 477 478 21 Karen JK, Gareau DS, Dusza SW, Tudisco M, Rajadhyaksha M, Nehal KS. Detection of basal 479 cell carcinomas in Mohs excisions with fluorescence confocal mosaicing microscopy. Br J 480 Dermatol 2009;160:1242–50. https://doi.org/10.1111/j.1365-2133.2009.09141.x. 481 482 Keena S, Que T. Research Techniques Made Simple: Noninvasive Imaging Technologies for 483 the Delineation of Basal Cell Carcinomas. J Invest Dermatol 2016;136:e33–8. 484 https://doi.org/10.1016/j.jid.2016.02.012. 485 486 Kose K, Bozkurt A, Alessi-Fox C, Brooks DH, Dy JG, Rajadhyaksha M, et al. Utilizing 487 Machine Learning for Image Quality Assessment for Reflectance Confocal Microscopy. J 488 Invest Dermatol 2020;140:1214–22. https://doi.org/10.1016/j.jid.2019.10.018. 489 490 Liang X, Wu L, Li J, Wang Y, Meng Q, Qin T, et al. R-Drop: Regularized Dropout for Neural 491 Networks 2021. arXiv:2106.14448 492 493 Mu EW, Lewin JM, Stevenson ML, Meehan SA, Carucci JA, Gareau DS. Use of digitally 494 stained multimodal confocal mosaic images to screen for nonmelanoma skin cancer. JAMA 495 Dermatology 2016;152:1335–41. https://doi.org/10.1001/jamadermatol.2016.2997. 496 497 Nasr-Esfahani E, Samavi S, Karimi N, Soroushmehr SMR, Jafari MH, Ward K, et al. Melanoma 498 detection by analysis of clinical images using convolutional neural network. Proc. Annu. Int. 499 Conf. IEEE Eng. Med. Biol. Soc. EMBS, vol. 2016- October, Institute of Electrical and 500 Electronics Engineers Inc.; 2016, p. 1373–6. https://doi.org/10.1109/EMBC.2016.7590963. 501 502 22 Nehal KS, Bichakjian CK. Update on keratinocyte carcinomas. N Engl J Med 2018;379:363–503 74. https://doi.org/10.1056/NEJMra1708701. 504 505 Leiter U, Eigentler T, Garbe C. Epidemiology of Skin Cancer. Sunlight, Vitam. D Ski. Cancer, 506 vol. 810, New York, NY: Springer New York; 2014, p. 120–40. https://doi.org/10.1007/978-1-507 4939-0437-2_7. 508 509 Pan SJ, Yang Q. A survey on transfer learning. IEEE Trans Knowl Data Eng 2010;22:1345–510 59. https://doi.org/10.1109/TKDE.2009.191. 511 512 Panarello D, Compérat E, Seyde O, Colau A, Terrone C, Guillonneau B. Atlas of Ex Vivo 513 Prostate Tissue and Cancer Images Using Confocal Laser Endomicroscopy: A Project for 514 Intraoperative Positive Surgical Margin Detection During Radical Prostatectomy. Eur Urol 515 Focus. 2020 Sep 15;6(5):941-958. doi: 10.1016/j.euf.2019.01.004 516 517 R Core Team (2020). A language and environment for statistical computing. R Foundation for 518 Statistical Computing, Vienna, Austria. n.d. 519 520 Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez JC, et al. pROC: An open-source 521 package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 2011;12. 522 https://doi.org/10.1186/1471-2105-12-77. 523 524 Rogers HW, Weinstock MA, Feldman SR, Coldiron BM. Incidence estimate of nonmelanoma 525 skin cancer (keratinocyte carcinomas) in the us population, 2012. JAMA Dermatology 526 2015;151:1081–6. https://doi.org/10.1001/jamadermatol.2015.1187. 527 23 Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual 528 Explanations from Deep Networks via Gradient-based Localization. Int J Comput Vis 529 2016;128:336–59. https://doi.org/10.1007/s11263-019-01228-7. 530 531 Schüürmann* M, Stecher MM, Paasch U, Simon JC, Grunewald S. Evaluation of digital 532 staining for ex vivo confocal laser scanning microscopy. J Eur Acad Dermatology Venereol 533 2019:jdv.16085. https://doi.org/10.1111/jdv.16085. 534 535 Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. 536 Nat Med 2019;25:44–56. https://doi.org/10.1038/s41591-018-0300-7. 537 538 Van Loo E, Mosterd K, Krekels GAM, Roozeboom MH, Ostertag JU, Dirksen CD, et al. 539 Surgical excision versus Mohs’ micrographic surgery for basal cell carcinoma of the face: A 540 randomised clinical trial with 10 year follow-up. Eur J Cancer 2014;50:3011–20. 541 https://doi.org/10.1016/j.ejca.2014.08.018 542 543 Wodzinski M, Skalski A, Witkowski A, Pellacani G, Ludzik J. Convolutional Neural Network 544 Approach to Classify Skin Lesions Using Reflectance Confocal Microscopy. Proc. Annu. Int. 545 Conf. IEEE Eng. Med. Biol. Soc. EMBS, vol. 2019, Institute of Electrical and Electronics 546 Engineers Inc.; 2019, p. 4754–7. https://doi.org/10.1109/EMBC.2019.8856731. 547 548 Wodzinski M, Pajak M, Skalski A, Witkowski A, Pellacani G, Ludzik J. Automatic Quality 549 Assessment of Reflectance Confocal Microscopy Mosaics using Attention-Based Deep Neural 550 Network. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, vol. 2020- July, Institute 551 24 of Electrical and Electronics Engineers Inc.; 2020, p. 1824–7. 552 https://doi.org/10.1109/EMBC44109.2020.9176557. 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 Supplementary Table 1. Summary of patient´s demographic data and lesion location Abbreviation: SD, standard deviation  Variable Total (53 patients/79 lesions) Age, mean (SD), y* 61 (13) Sex, %(n) Male 64% (34/53) Female 36% (19/53) Diagnosis, n(%) 79 lesions BCC 47 (59.5%) Nodular 18 (22.8%) Superficial 11 (13.9%) Infiltrative 10 (12.6%) Mix-subtype 8 (10.1%) “Not-BCC” 32 (40.5%) Location, n(%) 79 lesions Head and neck 50 (63.3%) Upper extremities 8 (10.1%) Trunk 15 (19%) Lower extremities 6 (7.6%) 1 SUPPLEMENTARY MATERIAL 1 Supplementary Figure 1. Algorithm´s performance in each separate category with three 2 different trainings. 3 4 Supplementary Figure 2. EVCM image acquisition: a) Purple and pink digitally colored 5 EVCM image of a nodular BCC (yellow boxed area and arrows) acquired from a freshly excised 6 whole-tissue obtained from Mohs surgery. EVCM images tissues simultaneously in 7 fluorescence and reflectance mode, where fluorescence signal from nucleus is converted to 8 purple color (b) and reflectance signal from cytoplasm and collagen is converted to pink color 9 (c) to generate a combined purple and pink image (d), resembling conventionally H&E- stained 10 image. Scale bars: a) 400 µm and b, c, d) 200 µm. 11 12 Supplementary Table 1. Summary of patient´s demographic data and lesion location 13 14 Supplementary Video 1. Steps involved in EVCM image acquisition. 15 16