High‐resolution full‐field optical coherence tomography microscope for the evaluation of freshly excised skin specimens during Mohs surgery: A feasibility study

Histopathology for tumor margin assessment is time‐consuming and expensive. High‐resolution full‐field optical coherence tomography (FF‐OCT) images fresh tissues rapidly at cellular resolution and potentially facilitates evaluation. Here, we define FF‐OCT features of normal and neoplastic skin lesions in fresh ex vivo tissues and assess its diagnostic accuracy for malignancies. For this, normal and neoplastic tissues were obtained from Mohs surgery, imaged using FF‐OCT, and their features were described. Two expert OCT readers conducted a blinded analysis to evaluate their diagnostic accuracies, using histopathology as the ground truth. A convolutional neural network was built to distinguish and outline normal structures and tumors. Of the 113 tissues imaged, 95 (84%) had a tumor (75 basal cell carcinomas [BCCs] and 17 squamous cell carcinomas [SCCs]). The average reader diagnostic accuracy was 88.1%, with a sensitivity of 93.7%, and a specificity of 58.3%. The artificial intelligence (AI) model achieved a diagnostic accuracy of 87.6 ± 5.9%, sensitivity of 93.2 ± 2.1%, and specificity of 81.2 ± 9.2%. A mean intersection‐over‐union of 60.3 ± 10.1% was achieved when delineating the nodular BCC from normal structures. Limitation of the study was the small sample size for all tumors, especially SCCs. However, based on our preliminary results, we envision FF‐OCT to rapidly image fresh tissues, facilitating surgical margin assessment. AI algorithms can aid in automated tumor detection, enabling widespread adoption of this technique.


| INTRODUCTION
Nonmelanocytic skin cancer (NMSC) is the most prevalent cancer worldwide, accounting for $5.4 million cases diagnosed and treated annually in the United States alone [1].Among all NMSCs, basal cell carcinoma (BCC) is the most common type ($4.3 million cases), followed by squamous cell carcinoma (SCC; $1 million cases) [1].NMSCs are rarely fatal and seldom metastatic, but they can be highly infiltrative and aggressive and have a high recurrence rate [2].
Surgical excision and Mohs micrographic surgery are widely accepted procedures for the margin assessment and complete removal of the NMSC with a high cure rate of 95%-99%, respectively [3].To achieve a high cure rate and preserve healthy skin, histopathological examination of the excised tissue is the gold standard.However, histopathology evaluations require timeconsuming tissue preparation, extensive laboratory facilities, and well-trained technicians [4,5].
Ex vivo optical imaging devices, including confocal microscopes and optical coherence tomography (OCT), have been developed for the rapid evaluation of fresh tissues to obviate tissue processing [5][6][7][8][9][10][11].In this article, we describe the utility of a novel full-field OCT (FF-OCT) microscope (ApolloVue ® B100 image system, Apollo Medical Optics, Ltd.) device.OCT relies on a lowcoherence interferometer and the light-scattering properties of skin structures to construct cross-sectional images of tissue [12].However, the existing OCT devices have a low resolution (3-10 μm axial and 3-7.5 μm lateral resolution), which hinders the differentiation of normal skin structures from tumors and tumor subtyping [12,13].In contrast, the novel FF-OCT microscope has an axial resolution of 1.5 μm and a lateral resolution of 1.1 μm, which is far superior to the existing OCT devices.Wang et al. demonstrated that even a novice (without OCT experience) can read these images with 93%-100% sensitivity and 21%-54% specificity.However, this study was performed on paraffin-embedded thick tissue sections, which does not equate to the evaluation of freshly excised tissues [14].
Although the emergent cellular-resolution OCT could significantly accelerate the clinical adoption of OCT to assist physicians in interpreting images [15][16][17][18], interpretation of OCT images often requires an expert with extensive training in reading these images, posing a major barrier to integrating OCT in clinics [6,9].Thus, deep learning algorithms, in particular convolutional neural networks (CNNs), have become a powerful tool for analyzing medical images to assist physicians in detecting, classifying, segmenting, and even diagnosing tissue images [19].CNN has the advantage of automatically extracting features and is not limited to features defined by the human eye.At present, many studies have used CNN to identify basal cell carcinoma in stained images [20] and segment nuclei from stained images [21,22] and dermal fillers in OCT images of mouse skin [23].
In our study, for the first time, we imaged fresh, non-labeled (without using any exogenous dye/contrast agent) tissues obtained during Mohs surgery using a high-resolution FF-OCT microscope.First, we defined FF-OCT features of all normal skin structures and various NMSC tumors, particularly BCC, and some benign lesions.Next, to demonstrate the feasibility of this device in a surgical setting, we performed a blinded analysis by two OCT experts (one pathologist and one Mohs surgeon) to access the diagnostic accuracy (sensitivity and specificity) of detecting tumors.Finally, to overcome the limitations of reading grayscale images and mitigate the mosaic artifacts due to image stitching, we generated a deeplearning algorithm that can differentiate BCC tumor nodules from sebaceous glands and other nontumor tissues.

| Patient cohort, consenting, and tissue collection
This study was conducted at Memorial Sloan Kettering Skin Cancer Center, Hauppauge, New York between April 2017 and February 2020.Patients undergoing Mohs surgery for NMSCs consented to an institutional review board-approved protocol (#08-006) for the collection of fresh discarded specimens (cut tangentially) after the completion of pathology analysis.Our collection did not compromise Mohs' procedures or patient care.We excluded patients younger than 18 years of age.Also, large samples that exceeded the imaging window area (20 Â 20 mm 2 ) of the tissue holder or tissues thinner than 1 mm were not included in the study.

| Tissue preparation and FF-OCT imaging
Discarded tissues were thawed from the frozen blocks and rinsed in normal saline.They were then placed in the plastic cassette (provided with the device) with the frozen section cut surface facing the glass window for imaging.Nicks and their associated color codes were applied on the edges of the specimens for the purpose of tissue orientation and subsequent histopathology correlation.A drop of glycerin was applied to the glass window before the tissue placement.The cassette was then closed with a cover to secure the specimen in place.The tissue flattening was achieved using the sponge cushion lining the inner surface of the cover.A drop of mineral oil was added to the lens and the cassette was then inserted into the imaging well of the device for scanning.The technological details of the device have been previously described [14].Once the scanning process was completed, the entire plane of 10-30 μm below the cut surface of the specimen was visualized as one mosaic composed of multiple small fields of view (FOVs; 800 μm Â 600 μm).The scanned images were stored on a connected computer for analysis [14].

| Blinded analysis
The two expert readers (MJ, a pathologist; and CSJC, a Mohs surgeon), first trained themselves by studying 10 mosaics from BCC tumors and defined FF-OCT features for BCC and surrounding normal structures (epidermis, hair follicles, sebaceous glands, eccrine ducts, adipose tissue, vessels, and nerves), comparing them with their corresponding histopathology.Later, a test set was created using 113 FF-OCT mosaics (from 113 fresh tissues).The images used for training were removed from the test set.All the FF-OCT images collected were deidentified and were assigned a study number and provided to the experts for reading independently.The readers were blinded to the histopathology diagnosis.Each of the readers recorded findings, including the presence or absence of the tumor, type, and subtype of tumor, in a spreadsheet.Clinical data was also collected for the consented patients including name, age, gender, clinical diagnosis, and location of the lesion.Corresponding histopathology sections (created at the time of Mohs evaluation) provided the closest mirror images of the FF-OCT mosaics and were used as ground truth for the concordance of FF-OCT reading.

| Deep-learning algorithm
For the artificial intelligence (AI) algorithm, images from 23 nodular BCC (nBCC) were used.The image sizes were variable, and each image had more than 5000 Â 5000 pixels with a pixel separation of 1.332 μm.The images were chopped into patches with 512 Â 512 pixels to accommodate the limitations of computation power and storage capacity.
A CNN classification model was built on top of the U-Net with symmetric down-and upsampling results for nBCC detection [17,24,25] (Figure S1).During the CNN training phase, 1253 image patches with nBCC were used.The largest receptive field is 186 Â 186 μm 2 .During training, the cross-entropy loss was used as the baseline for evaluation with fivefold cross-validation.In addition to mitigation of imbalanced nBCC and non-nBCC classes, the focal loss was also adopted to improve the segmentation performance.As shown in Equation ( 1), the focal loss is defined to down-weight easy examples and focus training on hard negatives.The focal loss downweights easy examples with a factor of (1 À pt)γ so that the model can focus on learning the misclassified pixels [24].
where α t is the weighting factor to address the class imbalance issue, γ is the focusing parameter, and p t is defined in Equation ( 2)].
where p [0,1], is the model's estimated probability for the class with y = 1 (nBCC pixels).In our two-class scenario, y = À1 for non-nBCC pixel.
The focal loss was first employed to eliminate the OCT mosaic artifacts.After the segmentation model, a classification model was used to differentiate the BCC tumor nodules from other tissues.
To quantitatively evaluate the image segmentation performance, mean intersection-over-union (mIOU) was used to measure the overlapping between the predicted and annotated image pixels.
In addition to image segmentation, a classification model was built on top of the U-Net result for nBCC detection of the excised tissues.The post-segmentation image erosion process was applied to reduce the fragmented dusty pixels.In addition, a voting strategy was adopted on the outputs of the U-Net patches by partitioning each of the 512 Â 512-pixel patches into 128 Â 128-pixel patches for both hard and soft votings.As a result, 22 386 small patches were generated for training, and among them, 10 193 small patches have nBCC pixels.Resnet18 was used as the classification model.And fivefold cross-validation was applied.

| Patient demographics and lesion site
One hundred ten patients were enrolled in this study.The male:female ratio was 1:4 and the average age of the patients was 63 years old (ranged 33-93 years).Majority of the lesions 54/113 (47.8%) were located in the T-zone region (ear, eye, nose, lip, and chin), followed by 39/113 (34.5%) in the head and neck region (scalp, forehead, cheek, and neck), 14/113 (12.4%) on the extremities (arms, legs, hands, and feet), and only 6/113 (5.3%) on trunk and genitalia.

| Normal skin structures
On a low magnification view, all skin layers, including epidermis, dermis, and subcutis, could be identified (Figure 1).On a zooming-in, cellular details of each layer became evident.The epidermis appeared as a grayish linear stratified layer composed of multiple cells with a small dark nucleus and grayish cytoplasm.However, due to the difficulties in flattening the tissue edge completely on the imaging window, the epidermis could not be visualized in most of the tissues.
Hair follicles (Figure S2) appeared as tubular to round structures with a central dark hole lined by an inner grayish epidermal layer and an outer bright fibrous layer.Sometimes, a bright hair shaft was identified in the center of these follicles making their identification easy.Sebaceous glands (Figure S2) appeared as round to oval varied-sized darkish (hypo-reflective) structures composed of multiple lobules separated by thin bright (hyper-reflective) fibrous septa.Due to their round shape, these glands were difficult to distinguish from nBCC; however, the presence of multiple small bright punctate particles, which we speculate to be the sebum particulates, aided in the distinction.Eccrine glands (Figure S2) appeared as tightly packed of small, round to oval grayish (hypo-reflective) structures separated by thin bright septa.Within the gland's lumen, small punctate bright particles (similar to sebaceous glands) could be seen.Eccrine ducts (Figure S2) were seen within the clusters of eccrine glands as small roundish structures with a central dark (areflective lumen) and lined by grayish cells.The eccrine unit (glands and ducts) could be identified as embedded within dark (areflective) adipose tissue.
Smooth muscles (Figure S3) could be identified as bundles of grayish (hypo-reflective) structures with intervening bright thin fibrous bands.Cigar-shaped dark elongated nuclei were seen within the muscle fibers.The smooth muscle bundles were seen lining the dark lumen of medium-sized blood vessels and attached to a hair follicle (as arrector pili muscle).

| Basal cell carcinoma
Classic features of BCC could be identified on FF-OCT.BCC tumor nodules appeared as round to oval varied size structures composed of clusters of grayish (hyporeflective) pleomorphic cells with dark nuclei (Figures 2 and  3).These tumor nuclei were seen arranged perpendicular T A B L E 1 Classification accuracies of two expert readers for detecting tumors in the surgically excised fresh tissue.at the periphery of the nodule forming "palisading."

Assessment
Clefting was identified as a dark (areflective area) around tumor nodules.
In sBCC, the tumor nodules were seen attached to the epidermis (Figure 3A).In nBCC (Figure 2) and mnBCC (Figure 3B), the nodules were identified within the bright dermis.Palisading and were prominent in both nBCC and mnBCCs.Necrosis was seen in the bigger tumor nodules of nBCC as dissociated cells with some scattered bright particles within.The collagen appeared as bright (hyper-reflective) parallel bundles around the tumor nodules.iBCC had a distinct appearance (Figure 3C).The tumor foci appear as darkish irregular strands (varied size and shape) composed of grayish clusters of cells with intervening bright strands of fibrous tissue.It was easier to identify these strands when they were clustered.Isolated foci were not readily detected.Additionally, in the area of iBCC, there was a complete loss of normal skin structures.

| Squamous cell carcinoma
SCC was seen as sheets of polygonal cells with abundant grayish cytoplasm and enlarged irregular dark nuclei.
Within the nucleus, a bright dot was often identified, which could be the nucleolus (Figure S4).

| Other tumors
Cylindroma (Figure S5) appeared as well circumscribed large multi-lobated grayish structure within the dermis.Some of the lobules were surrounded by a bright thickened band of collagen.Each lobule is composed of monomorphic cells with a dark round nucleus and a rim of scant grayish cytoplasm.No clefting around the nodule was seen.

| Diagnostic accuracy of FF-OCT device in detecting residual tumors in the surgically excised fresh tissue
Readers 1 (CSJC) and 2 (MJ) each demonstrated high sensitivity (91.6% and 95.8%) and moderate specificity (55.6% and 61.1%), respectively, for detecting the presence of any malignant tumor in the margin (Table 1).A fair degree of agreement was shown on this task, with Cohen's kappa = 0.327 (Table S2).Readers were less accurate in differentiating SCC from BCC; each reader indicated the presence of SCC in 7 out of 17 specimens containing a histologic SCC component.In addition, raters overcalled the presence of nodular BCC; they indicated the presence of this subtype in 54% and 58% of all cases and those instances were correct 55.7% and 60.0% of the time (positive values).

| Deep-learning algorithm
Most nBCC regions were segmented (Figure 4 and Table 2).The integrated segmentation and classification model showed better performance than the segmentation-only model.The artifact of the FOV boundary was significantly reduced when comparing the results of focal loss and crossentropy loss.The false positives on nBCC are significantly reduced, and the mIOU increased to 60.3 ± 10.1%.The sensitivity and specificity reach 93.5 ± 2.2% and 81.2 ± 9.2%, respectively.

| DISCUSSION
In this manuscript, we described, for the first time, cellular features of normal skin and NMSCs including BCC, and a benign lesion with a novel high-resolution FF-OCT device in fresh ex vivo tissues.Through a blinded analysis, we demonstrated the potential utility of this device for identifying and classifying neoplastic keratinocytic lesions.The FF-OCT had high sensitivity in detecting all the tumors but had a low-moderate specificity.
Among all tumor types, nBCC had the highest sensitivity for tumor detection, however, had a moderate specificity.False-positive results were due to the inability to distinguish tumor nodules from sebaceous glands.However, sebaceous glands exhibit bright punctate dots and have sparse surrounding collagen that can aid in differentiating their identification.Similar particles have been defined with other optical imaging techniques such as confocal microscopy [26] and Coherent Anti-Stokes Raman Scattering Microscopy [27].The deep learning algorithms employed could largely segment nBCC regions.However, there were still some false positives on nBCC segmentation, which might be due to the similar OCT appearance between nBCC and sebaceous glands.On the contrary, SCC and sBCC had low sensitivity but very high specificity.The low sensitivity could be a result of incomplete visualization of the epidermis, which hindered the detection of these tumors originating from the epidermis.Incomplete visualization of the epidermis was caused by the use of tangentially excised Mohs specimens and resultant tissue flattening issues encountered during imaging.We believe such an issue can be improved with a vertically excised specimen.In the future, it may also be possible to resolve the flattening issue by using the newly described digital tissue flattening [28], which can then expand the application of this device to evaluate Mohs surgical margins.An additional cause for less accurate results, especially for SCC, could be the lack of SCC images in the training data set.Similarly, iBCCs and mnBCCs had a low sensitivity.This could possibly be due to the difficulty of identifying small strands or foci of these tumors among the bright and dense collagenous background.
The inter-rater reliability (Cohen's kappa, Table S2) values of the two OCT experts were below 0.4.This could be related to the inability to differentiate BCC tumor nests from normal sebaceous glands or follicular epithelium by the grayscale imaging, especially for small BCC strands or nests.Ex vivo confocal microscope can create digitally colored purple and pink images that simulate hematoxylin and eosin-stained tissue sections.The major limitation of this study was a small sample of tumors.Another limitation is the grayscale nature of images that requires interpretation by experts in this field.Thus, future studies are warranted using a large sample size (including benign lesions) and performing a multi-reader diagnostic accuracy study.Furthermore, deep learning algorithms can be integrated to convert grayscale images into digitally colored purple and pink images [29], similar to the images created by an ex vivo confocal microscope.This would improve visualization of the OCT images and reduce the learning curve.Moreover, AI can aid in the automated detection of tumors, leading to its wider adoption [20,30].This work is a preliminary trial to see the feasibility of cellular-resolution OCT on segmentation/classification of ex-vivo human cancerous tissues.The present dataset is not large enough for a countable analysis of the explainability of the AI model.But, our earlier work on a mouse SCC model did have a large dataset ($1 TB) [31].How the feature extraction progresses, especially how the cellular-level features are captured, can be well explained through the heat maps at multiple stages of the classifier (from a few micron receptive field to a few 100-μm receptive field), making our deep learning algorithm interpretable.The analysis shows that cellular-level features are critical to the success of SCC detection.
It should also be noted that the image data complexity of the present human BCC tissues is far more diversified than that of the mouse SCC model.So, a larger data set is needed to have a countable explainability discussion.
Based on our pilot study, we envision FF-OCT as an alternative for time-consuming and tedious histopathology to enable a rapid assessment of the tumor margins in the surgical excision samples, potentially reducing their recurrence rate.At least, FF-OCT may offer a role in initial specimen screening for the margin status in the operation room to facilitate completeness of tumor removal before conventional histological confirmation.It can also be combined with the in-vivo imaging techniques such as OCT and confocal microscopy that have limited penetration depth and often cannot used to evaluate deeper surgical margins.Additionally, FF-OCT can analyze small biopsies at the bedside before they are submerged in formalin for further processing.If indicated, all or part of the specimen can be preserved for molecular analysis as the tissue is neither processed nor sectioned.Finally, since the FF-OCT images are digitally stored, they can be read and analyzed remotely by a specialist, as a telehealth tool [32], for evaluation of ex vivo tissue, especially beneficial for rural or underserved areas.Although different ex vivo imaging technologies exist, knowledge of this novel device is essential to the consumers so they can tailor their needs based on the device's cost and capability.
Ultimately, ex-vivo OCT may not necessarily replace the current rapid pathology process but may help to fill the gap in the under-served community or rural area where an extensive lab set-up and trained technicians may not be readily available.

AUTHOR CONTRIBUTIONS
Each of the co-authors acknowledges their participation in conducting the research leading to this manuscript and has agreed to its submission to be considered for publication.All the authors have read and approved the final manuscript.Manu Jain, Kiran Singh, and Chih-Shan Jason Chen performed the image acquisition and analysis.Manu Jain and Chih-Shan Jason Chen designed the study.Shu-Wen Chang, Sheng-Lung Huang, and Homer H. Chen performed the AI analysis.Nicholas R. Kurtansky performed the statistical analysis.Manu Jain and Chih-Shan Jason Chen drafted and revised the paper.
Images of normal layers of the skin.(A) FF-OCT image showing grayish epidermal lining (pink arrow), dermis (green asterisk), and underlying subcutis (pink asterisk).(B) A stratified layer of the epidermis (pink arrow) shows multiple round to oval dark nuclei surrounded by bright (grayish) cytoplasm.The underlying dermis appears grayish (green asterisk).(C) Adipocytes of the subcutis appear as dark polygonal structures separated by thin white septa (pink asterisk).(D-F) Corresponding histopathology section stained with Toluidine blue.Magnifications: (A) = 5.2 mm Â 5.2 mm; (B, C) = 300 Â 150 mm 2 ; (D) = 10Â; (E, F) = 40Â.FF-OCT, full-field optical coherence tomography.F I G U R E 2 Features of a nodular basal cell carcinoma (BCC) as compared to the normal adnexal structures.(A) FF-OCT image shows a big BCC nodule (yellow asterisk) surrounded by normal skin structures.(B) Zoomed-in area from the red box in image (A) shows cellular features of BCC with nuclear atypia, peripheral palisading (green arrow), and clefting (yellow arrow).The nodule is surrounded by a thick bright band of collagen (red arrows).Sebaceous glands (orange asterisks) have a lobulated appearance with bright punctate structures.The hair follicles (pink arrows) appear round to oval with a central lumen and epidermal lining.(C) Corresponding H&E-stained histopathology image.Scale bars = 1 mm.FF-OCT, full-field optical coherence tomography; H&E, hematoxylin and eosin.F I G U R E 3 Images of basal cell carcinoma (BCC) subtypes.FF-OCT images: (A) superficial BCC with grayish tumor nodules (yellow arrow).Epidermal attachment is not visible due to the lack of epidermis caused by inadequate tissue flattening.(B) Micronodular BCC with small grayish tumor nodules, and (C) infiltrative BCC with grayish tumor strands (yellow arrows).BCC nodules are surrounded by a collagenous matrix.(D-F) Corresponding histopathology sections with Toluidine blue.Magnifications: (A-C) = 2 Â 1.5 mm 2 ; (D-F) = 10Â.FF-OCT, full-field optical coherence tomography.

F
I G U R E 4 The image segmentation results of three patients.The 49 white color in the annotation column represents the nBCC regions.The scale bars are all 2 mm.The 50 color-coding of yellow, pink, cyan, and black represent true positive, false positive, false negative, and 51 true negative, respectively.CEL, cross-entropy loss, FL, focal loss.
Summary of the nBCC segmentation and detection results.
a Represent best results.