Non-Melanoma Skin Cancers Coverage from Every Angle

Developing a Digitized Deep Learning Model to Detect Basal Cell Carcinoma

By: Jenna Carter, PhD
Posted: Tuesday, May 11, 2021

An article published in Experimental Dermatology presented evaluations of a deep learning model developed for the automated detection of basal cell carcinoma. Gertruud A.M. Krekels, MD, PhD, of Eindhoven University of Technology, The Netherlands, and colleagues conducted this study using digitized whole-slide image files collected after Mohs micrographic surgery. They found that quantitative evaluation of whole-slide image files, compared with conventional glass slide assessments, may reduce the workload of histologic analysis.

“Deep learning is a machine learning technique that trains a model to perform classification tasks directly from images…[and] learn to detect different features…using multiple hidden layers,” stated Dr. Krekels and colleagues.

For this study, two deep learning models were developed based on 171 digitized hematoxylin and eosin frozen slides from 70 different patients. The first model had a U-Net architecture and was used for the segmentation of basal cell carcinoma; the second model had a convolutional neural network and was used to classify the whole slide as basal cell carcinoma or basal cell carcinoma–negative. Model performance was then measured by comparing model prediction with manual pixel-level ground truth. Additional quantitative results were obtained by calculating the Dice score and area under the receiver operating curve (ROC) between the automatic segmentation and ground truth within the manual annotation bounding box. (The Dice score is a commonly used metric for evaluating segmentation tasks in medical imaging.)

Findings revealed that the segmentation model yielded an average Dice score of 0.66 and an average area under the ROC of 0.90. The whole-slide classification model also showed an average area under the ROC of 0.90. According to Dr. Krekels and colleagues, digitized slides may help reduce the time-consuming procedure and cost burden of the traditional glass-slide histology methods.

Disclosure: The authors reported no conflicts of interest.

By continuing to browse this site you permit us and our partners to place identification cookies on your browser and agree to our use of cookies to identify you for marketing. Read our Privacy Policy to learn more.