CLASSIFICATION OF BASAL CELL CARCINOMA IN EX VIVO CONFOCAL MICROSCOPY IMAGES FROM FRESHLY EXCISED TISSUES USING A DEEP LEARNING ALGORITHM

Abstract

Ex vivo confocal microscopy (EVCM) generates digitally colored purple-pink images similar to H&E, without time-consuming tissue processing. It can be used during Mohs surgery for rapid detection of basal cell carcinoma (BCC); however, reading EVCM images requires specialized training. An automated approach using a Deep Learning algorithm to BCC detection in EVCM images can aid in diagnosis. 40 BCCs and 28 negative (“not-BCC”) samples were collected at Memorial Sloan Kettering Cancer Center to create three training datasets: 1) EVCM image dataset (663 images), 2) H&E image dataset (516 images), and 3) a combination of the two datasets. 7 BCCs and 4 negative samples were collected to create a EVCM test dataset (107 images). The model trained with the EVCM dataset achieved 92% diagnostic accuracy, similar to the H&E model (93%). The area under ROC was 0.94, 0.95, and 0.94 for EVCM, H&E, and combination trained models, respectively. We developed an algorithm for automatic BCC detection in EVCM images (comparable accuracy to dermatologists). This approach could be used to assist with BCC detection during Mohs surgery. Furthermore, we found that a model trained with only H&E images (which are more available than EVCM images) can accurately detect BCC in EVCM images.

INTRODUCTION

Basal cell carcinoma (BCC) is the most common skin cancer accounting for ~2 millions of cases annually in the United States alone [Rogers et al. 2015]. Biopsy, followed by histopathology, is the gold standard for diagnosis and subtyping of BCCs (as aggressive or non-aggressive) for appropriate management — surgical treatment for aggressive tumors versus topical treatment (non-surgical) for less aggressive BCCs. For aggressive or recurrent tumors, especially those located in cosmetically sensitive sites such as the face, Mohs micrographic surgery (MMS) is the treatment of choice with high cure rates [Van Loo et al. 2014]. However, MMS is a tedious and time-consuming procedure as it involves careful removal of skin cancer layer-by-layer. Each excised layer undergoes frozen sectioning and microscopic evaluation, requiring up to 20-45 minutes. Often times multiple layers are removed to achieve complete tumor clearance; thus, the entire surgical procedure may last for several hours [Keena and Que 2016], increasing patient’s waiting time, complications, and cost of the procedure. Additionally, frozen sectioning can cause tissue destruction and create artefacts that may hinder the final diagnosis.

To expedite the surgical procedure, various ex vivo optical imaging devices have been developed [Bennàssar et al. 2013; Dalimier and Salomon 2012; Gareau et al. 2009; Karen et al. 2009]. These devices can rapidly image freshly excised tissues at “near-histopathological” resolution, obviating the need for destructive and time-consuming tissue processing.

Ex vivo confocal microscopy (EVCM) is an emerging imaging technique that can evaluate freshly (un-processed) excised whole-tissue samples without the need for tissue processing (frozen sectioning). As there is no tissue processing involved, EVCM can image tissues rapidly (less than a minute for a tissue measuring up to 2 cm), reducing the time for the Mohs surgery and enabling real-time imaging in the surgical suite [Keena and Que 2016]. Furthermore, EVCM creates digitally colored purple and pink images (similar to H&E images) by converting fluorescence signal originating from the nucleus into purple color and reflectance signal from the cytoplasm into pink color. These digitally colored images can be read by Mohs surgeons trained in pathology [Mu et al. 2016]. Although EVCM has demonstrated an overall high sensitivity and specificity (~ 90%) for detection of BCC [Gareau et al. 2009; Karen et al. 2009] during MMS [Bennàssar et al. 2013], it is only been utilized in very few academic centers. We believe that the integration of an automatic algorithm for the detection of BCC in EVCM images could immensely aid Mohs surgeons, increasing adoption of this technology. Moreover, EVCM technology may be useful for surgical pathologists and dermatologists to obtain faster results from standard excisions [Bennàssar A et al. 2012; Debarbieux et al. 2015] or even for biopsies of inflammatory skin lesions [Bağcı et al. 2019; Bağcı et al. 2021; Bertoni et al. 2018]. This approach can achieve the goal of a real “bedside” pathology, similar to the ongoing integration of this technique for assessment of non-dermatology specimens in surgical pathology [Panarello et al. 2020].

Artificial intelligence (AI) is currently transforming healthcare [Hinton 2018]. A popular AI technique for image classification is Convolutional Neural Networks (CNNs), a deep-learning approach inspired from the human brain. In CNNs, inputs such as images go through several layers of artificial “neurons” before an output is finally rendered such as the diagnoses of those images. CNN algorithms are being used in radiology and pathology [Topol 2019] to classify images as neoplastic or non-neoplastic and have shown a proficiency at par or even exceeding human performance [Campanella et al. 2019]. Likewise, in dermatology, CNNs have reported comparable performance to an expert dermatologist in skin cancer diagnosis using clinical [Esteva et al. 2017; Fujisawa et al. 2019, Han et al. 2018] and dermoscopy images [Brinker et al. 2019; Codella et al. 2016; Haenssle et al. 2018; Haenssle et al. 2020]. Recently, CNN has also been applied successfully to reflectance confocal microscopy (RCM) images to classify skin lesions [Kose et al. 2020, Wodzkinski et al. 2019; Wodzkinski et al. 2020, Campanella et al. 2021]. However, to the best of our knowledge, CNNs have not been developed and tested to diagnose skin cancers in digitally colored EVCM images.

The goal of this study was to develop and test the performance of CNN algorithms for detecting BCC in digitally colored EVCM images obtained from freshly excised tissues from Mohs surgery. For this study, 40 BCCs and 28 negative (“not-BCC”) skin tissue samples were collected from 42 patients to create three different image datasets to train CNN models: 1) an EVCM image dataset with 663 images, 2) an H&E image dataset with 516 images, and 3) a combination of the two datasets (EVCM and H&E image datasets) with 1179 images. The performance of these 3 trained models were evaluated and compared on a separate test set (not used in training), which comprised of 97 EVCM images created using 7 BCCs and 4 negative (“not-BCC”) skin tissue samples were from 11 patients.

RESULTS

Patient demographics and lesion characteristics:

A total of 53 patients were enrolled in the study. Mean age was 61 years (± 13, range 36-95 years); 64% (34/53) were males and 36% (19/53) females. Majority of the lesions, 63.3% (50/79) were located on the head and neck. A total of 47 BCCs were imaged including 18 nodular BCCs (nBCCs), 11 superficial BCCs (sBCCs), 10 infiltrative BCCs (iBCCs), 8 mix-subtype BCCs and 32 skin samples did not present BCC. Patient demographics and lesion characteristics are detailed in the Supplementary Table S1.

Model performance:

The main outcome measures were sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) (Table 1 and Supplementary Figure S1).

Table 1.

Comparison of the performance of trained models for diagnosis of BCC in the EVCM images

Training Dataset	Results of BCC detection in EVCM images- Metrics (CI 95%)
	Sensitivity	Specificity	PPV	NPV	Accuracy
EVCM TRAINING	0.96 (0.88-1)	0.89 (0.80-0.97)	0.86 (0.75-0.97)	0.96 (0.91-1)	0.92 (0.86-0.98)
H&E TRAINING	0.93 (0.85-1)	0.92 (0.84-1)	0.89 (0.8-0.99)	0.95 (0.87-1)	0.93 (0.87-0.98)
EVCM + H&E TRAINING	0.78 (0.65-0.91)	0.92 (0.84-1)	0.87 (0.76-0.99)	0.85 (0.76-0.94)	0.86 (0.79-0.93)

Abbreviations: CI= confidence interval, PPV= positive predictive value, NPV =negative predictive value. Bold fonts indicate highest accuracy results.

The model trained with the EVCM dataset (EVCM model) achieved 92% diagnostic accuracy, similar to the H&E model (93%). Compared to the H&E model, the EVCM model had a higher sensitivity (96% vs. 93%) but lower specificity (89% vs. 92%). The combined model had the lowest diagnostic accuracy (86%) with a high specificity (92%; similar to H&E model) but the lowest sensitivity (78%). Area under the curve (AUC) of receiver operating characteristics (ROC) for diagnostic dichotomous classification was calculated for each the three training datasets (Figure 1). The AUC was 0.94, 0.95, and 0.94 for EVCM, H&E and combination trained models, respectively.

Figure 1. Receiver Operating Characteristic (ROC) Curves obtained using:

(a) EVCM-trained model, (b) H&E-trained model, and (c) combined (EVCM+H&E) trained model. The 95% CI bounds of the ROC curve were calculated via bootstrapping. The proposed algorithm achieves an AUC of 94.4%, 95.7% and 93.8%, respectively. The shaded ellipse represents the 95% CI area for the estimate of the sensitivity and specificity of the algorithm calculated via bootstrapping.

Gradient maps:

Gradient maps created with Grad-CAM [Selvaraju RR, 2016] highlighted important regions in red color in the images for predicting the presence/absence of tumor after all the three trainings. Figure 2 shows gradient map examples of a true positive (TP), a true negative (TN), a false positive (FP), and a false negative (FN) tissue sample. For the TP example, the algorithm identified even small BCC nodules (asterisks) as important areas for BCC prediction, while not taking the hair follicle (arrow) as an important region. Likewise, for the TN example, sebaceous gland was correctly identified as an important region within the image for the negative prediction for tumor. We had a very few images with false positive and false negative results. For FP example, sebaceous glands (arrowhead) and eccrine ducts (star) were detected as important for BCC detection. On the other hand, a FN example where BCC nodules (asterisks) were not considered important and the algorithm prediction was no-BCC.

Figure 2. Gradient Map: Left: EVCM images and Right: Gradient maps.

(a) True positive (TP) example of a BCC with small tumor nodules (asterisks). High prediction attributes over the tumor nodules in the Gradient map. Note, that a hair follicle (arrow) in the same field was not considered “important” for the prediction by the algorithm. (b) True negative example of an area with no-BCC, where a sebaceous gland (arrowhead) was identified “important” region for negative prediction. (c) False positive (FP) example where eccrine glands (star) and sebaceous gland (arrowhead) were detected as important for the prediction of tumor. (d) False Negative (FN) example, where BCC tumor nodules (asterisks) were not considered important by the algorithm for tumor detection. Note, this BCC tumor nodule has an extensive cystic degeneration in the center (dashed arrow), which could have resulted in the false prediction in this case. Color scalebar: Red color, high attribution and blue color low attribution for a given prediction by the algorithm. Scale bars: a, b, c, d) 200 μm.

DISCUSSION

BCC is the most common skin cancer worldwide [Leiter et al. 2014]. Although, BCC has a low metastatic potential it can be locally invasive causing extensive tissue damage and loss of regional function [Nehal and Bichakjian 2018]. MMS is a specialized surgical procedure capable of achieving complete clearance of BCC, while maximizing normal tissue preservation, making it a preferred treatment for recurrent BCCs and BCCs located on cosmetically sensitive and functionally challenging sites, such as the face [Jain et al. 2017]. However, MMS is not only a time-consuming surgery (due to frozen section analysis) but it is also an expensive procedure that requires an extensive laboratory set-up and specialized surgeons and technicians. Ex vivo confocal microscopy (EVCM) is an emerging imaging technique that generates digitally colored purple and pink images similar to H&E, without any time-consuming tissue processing [Mu et al. 2016; Schüürmann et al. 2019]. It can be used during Mohs surgery (MMS) for rapid detection of residual basal cell carcinoma (BCC); however, reading EVCM images requires specialized training. An automated approach to BCC detection in EVCM images can aid in diagnosis.

Currently, AI is being implemented extensively in the field of dermatology and pathology for the automated diagnosis of skin cancers and non-neoplastic lesions (psoriasis, atopic dermatitis and onychomycosis) [Han et al. 2018] in clinical and dermoscopy images using CNN, a deeplearning algorithm [Esteva et al. 2017, Schüürmann et al. 2019]. In our study we used CNN for detection of BCC in EVCM images. CNN was first used in skin cancer detection by Nasr-Esfahani et al. (2016) for the diagnosis of melanoma. They trained the algorithm with a small dataset of 170 clinical images from melanocytic lesions. Similar to our study, the authors used augmentation methods such as random rotation, and resizing of the images to increase the number of images (from 170 original images to 6120 images) in the training dataset, which yielded a sensitivity of 81% and a specificity of 80% in the diagnosis of melanoma.

CNNs has been applied to non-invasive in vivo imaging technique such as reflectance confocal microscopy (RCM). Wodzinski et al. (2019) reported an accuracy of 91% in the diagnosis of BCC in RCM images. Recently, Campanella et al. (2021) developed a deep learning model to automatically detect BCC in RCM images acquired from lesions clinically equivocal for BCC and compared the results with the RCM expert readers. The proposed model achieved an area under the curve (AUC) for the receiver operator characteristic (ROC) curve of 89.7%, which was on par with the expert readers. We achieved similar results with 92% diagnostic accuracy for detection of BCC in EVCM images using EVCM training dataset. The use of H&E images to train deep-learning algorithms has been extensively used in pathology. Towards this end, Campanella et al. (2019) reported an accuracy above 98% in the diagnosis of BCC in conventional H&E stained images. On the contrary, we achieved a diagnostic accuracy of 93% with H&E-trained model for the detection of BCC in EVCM images. This difference in diagnostic accuracy could be attributed to the differences in the type of images used for the training and testing datasets in our study i.e. H&E images trained model was tested on EVCM image dataset compared to the use of only H&E images in the training and testing datasets in their study. Another reason could be the use of a relatively larger number of images (9,962) used to train their model.

Our study demonstrates CNN’s high-level performance in classifying BCC in EVCM images. Even with the use of freshly discarded tumor margins in this study, which typically has less tumor burden than the central tumor de-bulk tissue, we achieved a high sensitivity and specificity in the diagnosis of BCC in EVCM images, which is at par with dermatologists’ reported level in the literature [Gareau et al. 2009; Karen et al. 2009; Mu et al. 2016]. The highest sensitivity value of 96% was obtained when the CNN was trained with EVCM images, whereas the specificity decreased to 89%. Our results showed the best accuracy with H&E-trained model, which could be attributed to the better and sharper image quality of H&E images compared to digitally-colored EVCM images. Because we demonstrated that H&E images can successfully train CNNs to diagnose BCC in EVCM images, one can collect a large dataset comprising of only H&E images to train such a model (as H&E images are more readily available than EVCM images). Due to large number of images used in the H&E training dataset, less time-consuming weakly supervised CNN training could be used [Campanella et al. 2019] instead of the completely supervised training used in our study.

Although the combined (H&E and EVCM training dataset) model improved the specificity of BCC detection in EVCM images, it had the lowest diagnostic accuracy and sensitivity compared to the other two models (EVCM and H&E models). It is possible that the model trained with the combined dataset may not have had sufficient training time (i.e., number of epochs), which can be explored in future work.

Furthermore, the gradient map created in this study could be combined with the model prediction to aid surgeons in the real-time diagnosis and also as a teaching-training tool for novices [Campanella et al. 2021].

Our study had some limitations. First, while our dataset covered all the common subtypes of BCCs, it had an overall small size samples of each subtype, which could not represent all the morphological appearance of BCCs encountered in clinical practice. Also, this study did not include analysis on pigmented BCCs, which may be important to be tested in future studies. Secondly, the algorithm has yet to be tested in actual clinical practice. Also, for imaging with the EVCM device, we used freshly discarded surgical margins from Mohs surgery, which often has less tumor burden than the central de-bulk tissue. Even with the smaller tumor burden in these samples our algorithm achieved high diagnostic values; thus, we anticipate better results using de-bulk specimens with higher tumor burden. Lastly, all the images were acquired at a single institution, which does not account for the variability in staining protocols and tissue processing. Thus, our results should be further validated on EVCM images acquired from various centers (multi-centric study).

In conclusion, we present the results of a deep-learning algorithm in classifying BCC in EVCM images. The various models developed could diagnose BCC in digitally colored purple and pink EVCM images which was at par with reported dermatologists’ accuracy in the literature. Furthermore, we found that a model trained with only H&E images (which are more available than EVCM images) can accurately detect BCC in EVCM images. Training deep-learning technology with H&E images to diagnose EVCM images expands the possibility our approach to be generalized to diagnose a variety of skin lesions (neoplastic and non-neoplastic) in excised tissues. Ultimately, deep-learning models could be integrated in existing EVCM devices to aid Mohs surgeons in identifying BCCs automatically. Prospective and larger scale studies are needed to validate this technology in real clinical practice.

MATERIAL AND METHODS:

Tissue sample collection and image acquisition:

All the tissues used for creating the training and test datasets were collected at Memorial Sloan Kettering Cancer Center (NY, USA) under institutional review protocols (IRBs #17-078 and # 08-066) approved by Memorial Sloan Kettering Cancer Center Ethics Committee and after written, informed consent from patients.

a). Tissue sample collection and image acquisition for EVCM images:

Freshly discarded whole tissues samples (excised en face tumor margins) were collected consecutively from BCC cases undergoing Mohs surgery under a prospective IRB protocol (# 08-066). These samples were dipped in 0.6 mM acridine orange dye (a fluorescent dye) for 20 seconds and immediately placed on a commercial EVCM device (Vivascope 2500; Caliber ID, Rochester, NY, USA) for imaging [Jain et al. 2017]. The tissue was similarly oriented as for the frozen sectioning, which enabled an en face view of the entire tissue. Digitally-colored purple and pink images were acquired by combining signals from the fluorescent and reflectance channels, respectively (Supplementary Figure S2). First, we acquired an overview image, covering the entire tissue (measuring up to 2 cm in maximum diameter) section. Then we acquired multiple smaller sized zoomed-in images (within the overview image) of BCC tumors and normal surrounding skin structures such as sebaceous glands, hair follicles, epidermis, and eccrine ducts.

b). Tissue sample collection and image acquisition for H&E images:

Under another IRB approved protocol (#17-078), a Dataline search was performed to identify lesions with a histopathology confirmed diagnosis of BCC. We retrieved routine histopathology H&E stained slides from these lesions from the Department of Dermatopathology. The glass slides were then digitized using Aperio AT2 slide Scanner (Leica Biosystems, Nussloch, Germany) in the Dermatology research lab. Similar to EVCM image acquisition, we acquired multiple H&E images at various magnifications (2x, 8x, 10x) from BCC tumors and normal surrounding skin structures (sebaceous glands, hair follicles, epidermis, and eccrine ducts). On average, we acquired 24 images (ranging from 1 to 49 images) of varied sizes measuring ~1200x600 pixels to ~12000x12000 pixels per tissue sample using EVCM and H&E images.

Image Labelling:

The H&E and EVCM images were analyzed for the presence or absence of BCC by a dermatologist (MSM) and a pathologist specialized in optical imaging techniques (MJ). Each image was labelled as “BCC” and “not-BCC” and used to create training and test datasets (see below).

Dataset Creation:

Training Datasets:

1. EVCM training dataset: A total of 663 digitally stained purple and pink EVCM images (190 “BCC” images and 473 “not-BCC” images) were obtained from 14 fresh BCC tissues (5 nodular BCC, 4 infiltrative BCC, 3 superficial BCC, and 2 infiltrative-nodular BCCs) and 15 negative (“not-BCC”) controls normal skin tissues.

2. H&E training dataset: A total of 516 H&E images (170 “BCC” images and 346 “not-BCC”) were obtained from 26 H&E stained slides of BCC (11 nodular BCC, 4 infiltrative BCC, 6 superficial BCC, 3 infiltrative-nodular and 2 superficial-nodular BCC), and 13 negative (“not-BCC”) controls normal skin tissues.

3. Combined EVCM and H&E training dataset: A total of 1,179 images, which was created by combining all 516 images from the EVCM database and all 663 images from the H&E dataset.

Test Dataset:

The CNN models built using the above three training datasets were tested on a new set of 107 EVCM images (45 BCC and 62 normal images) obtained from 7 BCCs and 4 normal skin tissue samples that were not previously shown to the algorithm. Images artifacts were introduced for both training and testing sets to simulate real conditions. Composition of the training and test datasets are detailed in Table 2. Example images from the training and test datasets can be seen in Figure 3.

Table 2.

Samples and images used for creating training and test datasets.

Training datasets	N° of Tissue Samples and their diagnosis		Total N° of images/dataset	N° of BCC positive and negative images
1. EVCM TRAINING DATASET	14 BCCs	5 nBCC (35.7%) 4 iBCC (28.6%) 3 sBCC (21.4%) 2 inBCC (14.3%)	663	“BCC” 190 (28,7%)
				“not-BCC” 473 (71,3%)
	15 normal (“not-BCC”) skin samples
2. H&E TRAINING DATASET	26 BCCs	11 nBC (42.3%) 4 iBCC (15.4%) 6 sBCC (23.1%) 3 inBCC(l 1.5%) 2 snBCC (7.7%)	516	“BCC” 170 (32,9%)
				“not BCC” 346 (67,1%)
	13 normal (“not-BCC”) skin samples
3. COMBINED EVCM + H&E TRAINING DATASET (Datasets 1 and 2)	40 BCC and 28 normal (“not-BCC”) skin samples 1,179 images: 360 “BCC” images and 820 “not-BCC” images
Test-set	N° of Tissue Samples and their diagnosis		N° of images	N° of BCC positive and negative images
EVCM TEST SET	7 BCCs	2 sBCCs (28.6%) 2 nBCCs(28.6%) 2 iBCCs (28.6%) 1 snBCC(14.3%)	107	45 “BCC” (44,4%)
	4 normal (“not-BCC”) skin samples			62 “not-BCC”(53,6%)

Abbreviations: BCC= basal cell carcinoma, nBCC=nodular BCC, sBCC= superficial BCC, iBCC= infiltrative BCC, inBCC= infundibular BCC, snBCC= superficial and nodular BCC.

Figure 3. Images used to create training and test datasets:

EVCM images (upper and lower panel: Training and Test datasets): Purple and pink digitally colored EVCM images of: (a, i) nodular BCCs, (b, j) superficial BCC, (c, k) infiltrative BCC, and (d, l) normal (“not-BCC”) skin tissue with epidermis (arrow) and pilosebaceous gland (arrowhead). Conventional H&E-stained images (middle panel) of a: (e) a nodular BCC, (f) a superficial BCC, (g) an infiltrative BCC, and (h) normal (“not-BCC”) skin tissue with epidermis (arrow) and hair follicles (arrowheads). BCCs are shown with an asterisk. Scale bars: a, b, c, e, i, j) 100 μm and d, f, g, h, k) 200 μm.

CNN Architecture:

In this study, we trained and evaluated ResNet50 (Residual Neural Network) [He et al. 2016], a type of CNN, using both EVCM and H&E images. Unlike a standard CNN architecture, a ResNet architecture can handle greater number of hidden layers (higher model complexity), allowing for the extraction of more complex patterns and features. We used 181 hidden layers in our ResNet. We integrated Transfer Learning in our CNN to improve the efficacy of feature extraction. Transfer Learning stores knowledge gained from another problem and applies it to a different problem [Pan and Yang, 2010]. In this case, we used a CNN model pre-trained with the popular ImageNet dataset (which includes images from a large number of categories, including animal, plant, and objects but not medical images) and applied its knowledge to train our BCC-detection models.

The distribution of images labelled as “BCC” and “not-BCC” were imbalanced because we used different number of images per class within each training dataset (Table 2). Images in each class were augmented, increasing the total number of images to 10000 per class. To perform data-augmentation, synthetic copies were created by applying image transformation methods such as random rotation, shifting, and resizing [Wodzinski et al. 2020]. To speed up model training time and reduce model complexity, all input images were resized to 300x300 pixels [Wodzinski et al. 2019]. A total of 20 epochs (number of complete passes through a training dataset) were used to train each of the models. A schematic of the methodology is shown in Figure 4.

Figure 4. Experimental workflow:

a) Tissue acquisition, imaging and datasets generation, and b) CNN model used in this study. A ResNet50 of 181 hidden layers and pretrained with ImageNet was used on “BCC” and “not-BCC” labelled images. Abbreviations: BCC= Basal Cell Carcinoma.

The dataset was divided in three different subsets (training, validation and testing) in order to avoid the overfitting. The training and validation subsets were used to train the CNN and the testing subset was only used to see how well the model performs on unseen data. Besides, we also used Dropout Regularization [Liang et al. 2021] to reduce overfitting. Dropout works by probabilistically removing, or “dropping out,” inputs to a layer, which may be input variables in the data sample or activations from a previous layer. It has the effect of simulating a large number of networks with very different network structure and, in turn, making nodes in the network generally more robust to the inputs.

Gradient maps:

Gradient maps were created with Grad-CAM [Selvaraju RR et al. 2016]. Grad-CAM is a technique for visual explanation of CNNs that highlights the regions of the input that are “important” for the prediction in a particular CNN model. Grad-CAM determines the weight for each of CNN feature maps to compute the weighted sum of the activations and then up sampling the result to the image size to plot the original image with the heatmap, highlighting the important regions (red color) for the model prediction. A subset of gradient maps is shown in Figure 2.

Statistical analysis:

All statistical analyses was performed in R (v4.03) [R Core Team, 2020]. The ability to discriminate the classes (“BCC” or “not-BCC”) inferred by the model was used to generate ROC curves with 95% intervals, using the package “pROC” (version 1.17.0.1) [Robin et al. 2011]. ROC curves and 95% confidence intervals for sensitivity, specificity, positive predictive and negative predictive values, and accuracy measures were generated.

Abbreviations Used:

BCC

Basal Cell carcinoma

MMS

Mohs Micrographic Surgery

AI

Artificial Intelligence

EVCM

Ex Vivo Confocal Microscopy

ROC

Receiver Operator Characteristic

AUC

Area Under the Curve

CNN

Convolutional Neural Network

DATA AVAILABILITY

The deep learning model that was trained to generate the presented results and the code is available at https://github.com/boxyware/confocal-ex-vivo.