Can Smartphone Images Improve on Dermoscopy for Diagnosis of Skin Cancer?
Posted: Friday, October 15, 2021
Smartphone-captured images may impair the proper diagnosis of non-melanoma skin cancer compared with dermoscopic images, according to Eli O. David, MD, of Bar-Ilan University, Ramat-Gan, Israel, and colleagues. In fact, the prediction of malignancy using artificial intelligence was found to be less accurate and less sensitive when using smartphone versus dermoscopic images. These findings were published in the Journal of Cancer Research and Clinical Oncology.
“Physicians and patients should be aware of a possible decrease in sensitivity whenever diagnosing by nonstandardized smartphone teledermatology,” stated the study investigators. “The use of convolutional neural network analytics does not close the already known gap in face-to-face diagnostic accuracy between dermoscopy and smartphone photos.”
Dermoscopic and nondermoscopic smartphone images captured by dermatologists or general practitioners were randomly selected from two biopsy-validated databases, a prior clinical trial, and published editorial images. Dermoscopic or smartphone-captured images of histopathology-verified non-melanoma skin cancer (n = 132, 170, respectively) and benign skin lesions that were excised as suspected skin cancer (n = 33, 28, respectively) were compared using dual convolution neural network performance metrics. Metrics that predicted malignancy using the raw image and metrics that process sonification of the original images were combined to create a unified malignancy classifier.
Discriminative power and overall accuracy were found to be slightly higher with dermoscopic versus smartphone imaging, with an AUC of 0.911 and 0.821, respectively. The percent of overall accuracy for the prediction of malignancy appeared to be greater using dermoscopic (87.8%) than smartphone-captured images (74.8%, P < .005). Further, the percentage of correctly diagnosed malignancies was seemingly increased for dermoscopic versus smartphone-captured images (95.5% vs. 75.3%, P < .001). However, the percentage of correctly identified non-melanoma skin cancers was found to be similar across image types.
Disclosure: For full disclosures of the study authors, visit link.springer.com.
