Deep Learning on Basal Cell Carcinoma In Vivo Reflectance Confocal Microscopy Data

Abstract

Background: Extended skin malignancies of the head and neck region are among the most common cancer types and are associated with numerous diagnostic and therapeutical problems. The radical resection of skin cancer in the facial area often leads to severe functional and aesthetic impairment, and precise margin assessments can avoid the extensive safety margins. On the other hand, the complete removal of the cancer is essential to minimize the risk of recurrence. Reliable intraoperative assessments of the wound margins could overcome this discrepancy between minimal invasiveness and safety distance in the head and neck region. With the help of reflectance confocal laser microscopy (RCM), cells can be visualized in high resolution intraoperatively. The combination with deep learning and automated algorithms allows an investigator independent and objective interpretation of specific confocal imaging data. Therefore, we aimed to apply a deep learning algorithm to detect malignant areas in images obtained via in vivo confocal microscopy. We investigated basal cell carcinoma (BCC), as one of the most common entities with well-described in vivo RCM diagnostic criteria, within a preliminary feasibility study. Patients and Methods: We included 62 patients with histologically confirmed BCC in the head and neck region. All patients underwent in vivo confocal laser microscope scanning. Approximately 382 images with BCC structures could be obtained, annotated, and proceeded for further deep learning model training. Results: A sensitivity of 46% and a specificity of 85% in detecting BCC regions could be achieved using a convolutional neural network model (“MobileNet”). Conclusion: The preliminary results reveal the potential and limitations of the automated detection of BCC with in vivo RCM. Further studies with a larger number of cases are required to obtain better predictability.