Detecting Skin Cancer Using a Neural Network Algorithm
Posted: Thursday, January 16, 2020
According to research published in JAMA Dermatology, an algorithm operating on a region-based convolutional neural network was roughly as successful as dermatologists at identifying skin cancer and predicting the risk of malignancy. The algorithm had been trained to detect and identify skin cancer using 1,106,886 image crops containing possible lesions.
“Detection of cutaneous cancer on the face using deep-learning algorithms has been challenging because various anatomic structures create curves and shades that confuse the algorithm and can potentially lead to false-positive results,” noted Seung Seog Han, MD, PhD, of Dermatology Clinic, Seoul, and colleagues. “The results of the study suggest that the algorithm could localize and diagnose skin cancer without preselection of suspicious lesions by dermatologists.”
The diagnostic study included a validation set of 673 patients from whom 2,844 images of possible skin lesions were collected between January 1, 2010, and September 30, 2018. Among the patients, tumors were malignant in 185 patients and benign in 305 patients; 183 patients were tumor-free. A mean number of 4.2 images per patient was analyzed by the algorithm, which ordered them according to malignancy output. The algorithm was found to perform with both high sensitivity (76.8%) and high specificity (90.6%), with an area under the receiver operating characteristic curve of 0.910.
A test set of 325 images from 80 patients was also analyzed by human evaluators, including 13 board-certified dermatologists, 34 physicians performing a dermatology residency, 20 physicians in fields other than dermatology, and 52 people who had no professional medical background. The certified dermatologists and the algorithm had comparably accurate analyses, whereas the algorithm was found to be more accurate than the nondermatologic physicians, based on the findings of two statistical tests.
Disclosure: For full disclosures of the study authors, please visit jamanetwork.com.
