Melanoma is a skin cancer that kills nearly 10,000 people a year in the U.S. alone. As a result, it should not be a surprise that dermatology pops up often in our articles as a medical specialty that has been the target of digital health tech developments. We’ve written about artificial intelligence (AI) and neural nets used to diagnose skin cancer. Store and forward telemedicine can improve specialist access for dermatologists and 3D scanners enable teledermatology. UV-sensitive smart skin patches can tell people when to get out of the sun. These applications and more add to the dermatologist’s toolbox.
Researchers at the School of Computing, Mathematics, and Digital Technology, Manchester Metropolitan University (MMU), Manchester, have developed an AI system for automatic skin lesion region of interest (ROI) detection. Manu Goyal and Moi Hoon Yap of MMU trained two deep learning convolutional neural network (CNN) algorithms to detect ROI based on standard images taken with a smartphone. According to Goyal and Yap, there had been no earlier work with CNNs for ROI detection. The team tested their app against the current state-of-the-art dataset segmentation techniques. The results of the CNN testing were superior to segmentation in ROI detection.
Based on the ROI identification, the MMU researchers have also proposed a method to augment the data in images used to detect regions automatically. Traditional skin lesion dataset images have resolutions ranging from 540 x 722 to 4499 x 6748, while the images used with deep learning algorithms are typically much smaller, often 224 x 224 resolution. The MMU data-augmentation method uses multiple angles of the ROI without capturing redundant data or unwanted artifacts. Taking the ROI process a step further by adding mobility, Goyal and Yap designed the app to run on Samsung A5-powered smartphones using a MoleScope magnifying attachment.
The next steps for the MMU team include training the deep learning application for both localization and classification and adding the ability to work with many types of skin lesions. Research like this has the potential to make complex diagnostic skills readily available to the average person, even in under-served areas. This could help them decide when a skin lesion needs professional attention, and could save lives in the process.