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.

Overview

Regardless of complexion, everyone is at risk for skin cancers. Melanoma is the deadliest form with annual incidence rates increasing, resulting in an estimated 7,300 new cases and 9,000 deaths in the U.S. Melanomas are the most commonly diagnosed cancers among young people ages 14-20. Additionally, reoccurrence rates are high, with more than half of all patients developing new cancers in five years. Elevated recurrence may reflect current treatment options with excisional therapies being predominate, leading to a reticence to remove large margins of skin because of cosmetic concerns. Moreover, multiethnic individuals whose skin tans or is pigmented, underestimate their cancer risk. Inaccurate cancer risk perceptions lead to delayed skin cancer diagnoses. Although ethnic groups with darker skin have a lower incidence, ethnicity is a poor proxy for risk because melanomas are more prevalent and more frequently fatal in ethnic minority groups. Hawaii’s multiethnic populations, who experience year-round ultraviolet radiation (UVR) exposure and therefore higher risks for melanoma, provide a unique opportunity to identify ways to reduce the burden of this disease.

Objective/Aims

SMART is a prospective pilot feasibility study testing the preliminary effectiveness of a Digital Learning Computer Vision (DLCV) platform to distinguish melanomas from this scored set of biopsy-confirmed images. Our goal is to establish a DLCV method to triage lesions appropriate for biopsy; while providing a platform for increased vigilance of benign lesions. Our aims are to:

  1. Identify and label a set of de-identified images including a global, close and scope-based images of pigmented skin lesions, clinically diagnosed as melanoma and nonmelanoma from Hawaii Pathology Laboratories medical records to train the UH Cancer Center’s DLCV to identify melanocytic lesions.
  2. Compare and contrast the UH Cancer Center’s DLCV ability to distinguish melanomas versus nonmelanoma skin lesions using four digital image formats including 1) a dermoscopic image only; 2) a close image only; 3) a global and dermoscopic image, and 4) a global and close image. Our Hypothesis: A specific image format will be superior in discriminating melanomas from benign lesions.
  3. Validate the UH Cancer Center’s DLCV efficiency in distinguishing melanomas by comparing the results and timeliness of the diagnostics assessments made by a panel of dermatologists using the same image data. Our Hypothesis: The DLCV platform will be accurate and efficient in distinguishing legions suitable for biopsy. 

If successful, our DLCV platform will eliminate the need for excessive excisional biopsies and improve opportunities for dermatologists to provide higher levels of preventative care for at-risk patients.

Research Team

Kevin Cassell, DrPH

Principal Investigator

Christopher Lum, MD

Pathologist

Mark Willingham

Graduate Research Assistant

John Shepherd, PhD

Co-Investigator

Shane Spencer

Graduate Research Assistant