Skip to content

Shepherd Research Lab

    Systematic Melanoma Assessment and Risk Triaging Study (SMART)

    Our goal is to establish a deep learning computer vision (DLCV) method to triage lesions appropriate for biopsy while providing a platform for increased vigilance of benign lesions.


    Regardless of skin complexion or ethnicity, everyone is at risk for skin cancer. Skin cancer remains the most commonly diagnosed cancer, with more than 5 million new cases diagnosed in the United States (US) each year. Melanoma has a nearly 95% cure rate if treated early, so improved screening and monitoring programs for early detection will reduce morbidity, mortality, and costs.

    Dermoscopy is the standard clinical assessment method for skin cancers. Lesion morphology (size and shape) is qualitatively evaluated to select atypical skin lesions for biopsy. This approach can achieve 90% sensitivity, but only 59% specificity when performed by an expert. This poor diagnostic precision adds to the 1.6 billion dollars in annual therapeutic costs for melanoma.

    Artificial intelligence (AI) deep learning (DL) classification techniques have been shown to increase diagnostic sensitivity and specificity for melanoma. AI DL networks can highlight unique categorization information to develop a hypothesis on the underlying biology associated with lesion classification. However, these algorithms require further investigation into how it accurately classifies lesions and its use of benign and precursory nevi images. Furthermore, AI DL platforms have been trained and validated on predominately Caucasian and Asian populations. The accuracy of AI DL on African Americans, Hispanics, and Native Hawaiian/Pacific Islanders is unknown. Appropriate training of AI DL models using varied complexions and with varied types of skin lesions can improve generalizability to identify suspicious lesions early in high-risk multiethnic populations.

    Total body skin examination (TBSE) by dermatologists is another approach that allows for dermatologists to identify features and changes in lesion structures, leading to more accurate diagnoses. However, performing TBSE in the context of high-volume dermatological practices is difficult. Current clinical dermatological practice requires visual inspection of 50 to 100 potentially malignant lesions in each of 50 daily patients. This clinical burden for dermatologists to visually inspect over 2,500 lesions each day, limited the capacity for the sequential assessment and analysis of suspicious lesions.


    Our study will employ the use of a novel two-dimensional total body sequential examination imaging platform (2D-TBSE) which will capture both standard and dermoscopic images of participants’ skin lesions. We will work with our UH Cancer Center’s Office of Community outreach and engagement to use our 2D-TBSE on participants in the UH Cancer Center’s Skin √ Project. These volunteer local multi-complexion participants will provide the lesion images to test the feasibility of our candidate AI DL platforms to effectively use dermoscopic or standard images from the 2D-TBSE and train these AI DL platforms to be used in future clinical settings.

    Our long-term goal is to reduce the costs, morbidity, and deaths associated with skin cancers.

    The overall objective of our study is to identify integrative strategies using AI to improve the accuracy of medical professionals performing image-based screening and diagnosis.

    An example acquisition using the DermaGraphix IntelliStudio. (Image Source: Canfield Scientific, Inc.) Templates are using to scan multiple regions of the body. Suspicious lesions are imaged with an included dermascope. All lesions are auto detected.

    Our central hypothesis is that the incorporation of AI to help clinicians triage lesions at high risk for skin cancer can reduce time to treatment and costs. The rationale for the proposed research is that integrating effective AI technologies with novel imaging technologies has shown promise in reducing skin cancer morbidity and mortality.

    We plan to test our central hypothesis and, thereby, accomplish our overall objective by pursuing the following specific aims:

    Aim 1. Utilize a novel two-dimensional total body sequential examination imaging platform (2D-TBSE) capturing both standard and dermoscopic images to distinguish the best of several AI DL platforms to visually diagnose skin lesions and provide formative training and feasibility data for future studies. Hypothesis: Using 2D sequential whole-body imaging testing standard versus dermoscopic will identify the most effective format and provide the training data needed to support AI DL lesion classification.

    Aim 2. Assess the use of AI stacking techniques to determine if these combined AI platform approaches improve the overall accuracy of visual diagnoses of skin cancers when coordinated with an awareness of the individual AI training sets’ demographic compositions. Hypothesis: the use of multiple, coordinated AI visual diagnostic platforms will synergistically improve accuracy.

    The outcomes of the proposed work is to identify the changes in skin lesions that correlate to subsequent risk for skin cancers. The positive translational impact of these outcomes is that we can integrate visually image-identified skin cancer risk with other somatic and behavioral factors to prevent or screen for skin cancers.

    Research Team

    Kevin Cassel, DrPH

    Principal Investigator

    Christopher Lum, MD


    John Shepherd, PhD


    Mark Willingham

    Community Health Educator

    Funding Source(s)

    University of Hawaii Cancer Center Logo

    Cancer Center Support Grant
    11/01/2022 – 11/31/2023

    Research with SRL

    Willingham Jr ML, Spencer S, Lum CA, Sanchez JMN, Burnett T, Shepherd J, et al. The potential of using artificial intelligence to improve skin cancer diagnoses in Hawai'i's multiethnic population. Melanoma Research [Internet]. 2021 Dec 1;31(6):504–14. Cite