Material
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 (20x20 mm) 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 described14. Once the scanning process was completed, the entire plane of 10 to 30 microns below the cut surface of the specimen was visualized as one mosaic composed of multiple small fields of view (FOV; 800 µm x 600 µm). The scanned images were stored on a connected computer for analysis14.
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 de-identified 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 5,000 x 5,000 pixels with a pixel separation of 1.332 μm. The images were chopped into patches with 512 x 512 pixels to accommodate the limitations of computation power and storage capacity.
A convolutional neural network (CNN) classification model was built on top of the U-Net with symmetric down and up samplings result for nBCC detection17,24,25 (SupplementalFigure 1 ). During the CNN training phase, 1,253 image patches with nBCC were used. The largest receptive field is 186 x 186 μm2. During training, the cross-entropy loss was used as the baseline for evaluation with 5-fold 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 Eq. (1), the focal loss is defined to down-weight easy examples and focus training on hard negatives. The focal loss down weights easy examples with a factor of (1 − pt)γ so that the model can focus on learning the misclassified pixels24.
\(\text{FL}\left(p_{t}\right)=\ {-\alpha}_{t}\left(1-p_{t}\right)^{\gamma}log(p_{t})\)(1)
where \(\alpha_{t}\) is the weighting factor to address the class imbalance issue, γ is the focusing parameter, and \(p_{t}\) is defined in Eq. (2)].
\(p_{t}=\left\{\par \begin{matrix}p\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ ,if\ y=1\\ 1-p\ \ \ \ \ \ \ \ \ \ \ \ \ ,otherwise\\ \end{matrix}\right.\ \) (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= -1 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 512x512-pixel patches to 128x128-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 5-fold cross-validation was applied.