Research shows an increased risk of skin cancer among U.S. service members – leading the American Academy of Dermatology in 2018 to urge soldiers and veterans to work toward early detection.

Now, General Dynamics Information Technology has worked to develop an artificial intelligence-powered tool aimed at improving the skin care diagnostic process for veterans. The technology, as described on GDIT's website, is intended to classify skin lesions, determine if they are indicative of a common skin disease and potentially recommend follow-up care if necessary.

Dave Vennergrund, GDIT vice president of artificial intelligence and data insights, discussed the software with Healthcare IT News and talked through next steps for providers and patients.

Q. Could you tell me a bit about the partnership with the Department of Veterans Affairs?

A. As a participant in the 2021 VA National Artificial Intelligence Institute AI Tech Sprint, GDIT was poised to apply AI techniques to the question of how to better identify under-served women veterans. After sharing GDIT’s capabilities in image analytics, the VA pivoted our team to the challenge of diagnosing skin lesions.

VA NAII leader Dr. Rafael Fricks connected our team to Dr. Trilokraj Tejasvi. Tejasvi is a dermatologist in the Veterans Affairs Ann Arbor Healthcare System and clinical associate professor at University of Michigan.

Tejasvi was fundamentally interested in improving the process by which skin lesion images are captured by technicians and primary care physicians and forwarded to dermatologists and oncologists for preliminary diagnosis. He sought an analytical method that evaluates the clinical usefulness of an image to instantly identify poor images. Doing this rapidly would enable a new image to be captured at the initial consult, reducing potential delays.

Furthermore,  Tejasvi and our research team found that image classifiers built on deep learning models can be effective in classifying skin lesions into benign or malignant disease classes.

Q. How does GDIT’s tool work to help diagnose skin lesions?

A. The skin lesion classifier application enables a physician to easily upload a skin lesion image onto a web page and receive an instant classification and recommendation regarding follow-up care. The application was containerized for easy deployment into the primary care and hospital environments using cloud or on-premises hosting.

Q. What are the software’s current capabilities?

A. GDIT used transfer learning to leverage the image classification capability of open-source deep learning models: Resnet18, Resnet50, Resnext50, Densenet161, MobileNetV2, MixNet_Xl, EfficientNetB0, and EfficientNetB2. We trained the classifiers on 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).

The tool performed well when analyzing lesions on the training data publicly available. The highest accuracy as measured by the Accuracy Under the Curve was 0.92, produced by ResNext50, a variation on the powerful ResNet model.

Since all models displayed similar accuracy, and noting a severe imbalance in the training data, we concluded that training with more data will likely improve accuracy – independent of the specific deep learning model used.

In addition to unbalanced skin disease types, current assessments of skin lesion image repositories have shown that the images are unbalanced in another vital dimension – skin tones – with a preponderance of lighter skin tones 1 through 3 and a paucity of skin tones 4 through 6.

Q. What do next steps look like for the partnership and user access?

A. We have proposed a pilot with a series of steps for the next phase of development, including collecting more data from diverse skin tones and representing more disease types; training new classifiers on this extended data; and evaluating the solution for safety, utility and accuracy through standard clinical trial methods, including double-blind tests with VA oncologists, primary care providers and the veterans themselves.

The VA is leveraging numerous stakeholders from communities, including telehealth, diversity, health equity, innovation and AI research, to support the next phase of development and pilots.

Q.  How will GDIT ensure all skin tones are accurately represented and that bias will be avoided?

A. To be useful to a broader VA population, the training data representing all skin types needs to be collected and labeled – and then used to retrain the classifiers. Efforts across the VA are underway to build out a repository of such images.

Q. How will this complement VA’s telehealth offerings?

A. The current solution can be integrated into the VA telehealth constructs and frameworks. It is a fully containerized software solution that can run anywhere, on-site or in a cloud. It will specifically support the VA’s teledermatology practice and provide veterans the option to receive care as they choose, which is in alignment with the Mission Act.

The support for telehealth for dermatology was significantly lagging radiology, so this tool provides the opportunity to bring industry best practice in support of our veterans.

Q.  What are future implications of the project and software as a whole?

A. Skin cancer is the most common cancer in the United States, and one in five Americans will develop skin cancer in their lifetime. Risk can be even higher for veterans. Early screening and treatment are critical to improve a patient’s prognosis. The project has the potential to save thousands of veterans from skin cancer each year.

Kat Jercich is senior editor of Healthcare IT News.
Twitter: @kjercich
Email: kjercich@himss.org
Healthcare IT News is a HIMSS Media publication.