Skin cancer is the most common cancer in the U.S. with one in five Americans developing it in their lifetime. This risk can be even higher for service members and veterans – studies have shown they are often exposed to chemicals and high levels of ultraviolet radiation which can increase the risk of developing skin cancer. This makes the early screening and treatment even more critical to improve health outcomes.
Working together with the Department of Veterans Affairs, GDIT developed an accessible tool that will quickly identify if a lesion may be cancerous. It uses deep learning artificial intelligence (AI) to classify images of skin lesions into seven common categories, determine if an image is indicative of a common skin disease, and recommend immediate follow-up care.
The GDIT skin lesion classifier tool was initially developed during an AI tech sprint challenge organized by the VA. Together with Dr. Trilokraj Tejasvi, associate professor at University of Michigan Health and Chief of Dermatology at the VA Ann Arbor Healthcare System, the tool was developed to address the challenge of identifying skin cancer early the veteran community.
To do it, the team applied transfer learning techniques to deep learning models, comparing seven image classifiers built using deep learning algorithms. The classifiers were trained on a library of 30,000 publicly available skin lesion images, labeled with seven skin diseases: melanocytic nevi (benign), melanoma (malignant), benign keratosis-like lesions (benign) basal cell carcinoma (malignant), actinic keratoses and intraepithelial carcinoma (malignant), vascular lesions (benign) and dermatofibroma (benign).
From there, the team selected the most accurate model and built a cloud-based, containerized solution that could assess the images, identify a lesion as either malignant or benign, classify the condition, and make follow-up care recommendations. In addition, the team explored approaches to help explain the decision to evaluators by outlining the features and elements of an image that impacted its classification.
The skin lesion classifier tool is hosted in AWS, running in a container. It can be accessed through any web browser. A physician will be able to simply pull a skin lesion image onto a web page and receive an instant classification and recommendation regarding follow-up care.
Being able to reach veterans virtually is also paramount as many do not live near a VA medical center. Tejasvi is also the director of the University of Michigan Teledermatology Service and the teledermatology physician champion for the Veterans Integrated Service Network (VISN) 10. Even prior to the COVID-19 pandemic, the VA already had the largest telehealth program in the country. By mid-2020, VA telehealth visits increased by 1,000%.
Dave Vennergrund, GDIT vice president for AI and Data Insights, sees how AI solutions can improve the care veterans receive. “Imagine you could have a telehealth application that a patient could use and hold, and they could take a picture with their phone. This would then go through a workflow and if it was concerning it would set up an instant appointment through some scheduling mechanism to have the patient seen,” he said.
Today, a veteran with a skin lesion typically sees a Primary Care Provider (PCP) for a diagnosis. The PCP uses a camera to take an image of the lesion. The image is sent to dermatologists, centralized at 18 VISNs around the country, for review. Any issues with image quality are detected days later, and poor-quality images require the veteran to return for another image and start the whole process over again. Aside from being inconvenient and inefficient, this process may delay the start of potentially life-saving treatment.
The skin lesion classifier tool is being used in initial trials of a soft testing phase led by Tejasvi and Vennergrund, followed by seeking the approval to conduct clinical trials at the VA’s Ann Arbor hospital. To improve the tool, GDIT is working to increase the volume of data processed by the AI platform to ensure it captures as many unrepresented disease types as possible, including images of different skin tones.