Improving the Diagnosis of Melanoma and Precursor Lesions Among Veterans: Developing AI Techniques and Teledermatopathology
Abstract
The diagnosis of melanoma is among the most challenging areas of histopathology, because melanocytic lesions often demonstrate complex cellular and architectural features which must be evaluated as part of the diagnosis. While the gold standard diagnosis of melanocytic lesions is based on pathologists’ visual inspection of these features, our previous work has indicated a concerning level of diagnostic variability and error among pathologists. We have found that diagnoses of melanoma in situ and early stage invasive melanoma by actively practicing US pathologists are neither reproducible nor accurate. Achieving correct diagnoses is crucial for studying the process of transformation of a normal melanocytic lesion into melanoma. Achieving accurate diagnoses is also required in order for our Veterans to receive clinically appropriate treatment. The proposed research directly responds to the Fiscal Year 2019 Melanoma Research Program Focus Area of Precursor Lesions, Melanomagenesis, Host Factors, and the Tumor Microenvironment. Computer vision and machine learning techniques are powerful analytic tools that may improve pathologic diagnosis, allowing for computer systems to assess subtle features of hundreds of thousands of cells on a single biopsy simultaneously to identify diagnostic clues. While skin biopsies constitute a significant proportion of cases reviewed by pathology departments in the US Department of Veterans Affairs (VA) health care system, there is a lack of in-house dermatopathology expertise in VA facilities, which in turn inhibits the quality and efficiency of care provided to VA patients. Our study will support the establishment of a VA-wide teledermatopathology system to improve the quality of care provided to our VA patients while also developing a repository of whole slide digital images of skin biopsy cases for research (Aim 1). Our work will then leverage computer machine learning methods to improve accuracy in the diagnosis of melanoma and its precursors through evaluation of pathologists’ viewing behaviors as they interact with assessing these digital images (Aim 2) and the actual characteristics of the images (Aim 3). No substantial attempts have been made to understand errors in the diagnosis of melanoma or to evaluate possible solutions. Our study will identify underlying causes of diagnostic errors and guide future clinical and research efforts to improve the diagnosis of these challenging skin biopsy cases. Our goal is to implement strategies to reduce the burden of diagnostic errors on VA patients and our health care systems. The development of a VA teledermatopathology system will transform the ways in which scientific research and patient care is implemented in the VA system. The centralized data repository network that will be established will be a valuable scientific resource through which clinicians and researchers may collaborate on interesting and difficult cases, while serving as an operational mechanism through which in-house interpretations and subsequent diagnoses by dermatopathologists may be rendered. The proposed work will use computer vision and machine learning techniques to develop classifiers for region-of-interest detection and analysis; these are the beginning of future tools for computer-aided diagnosis.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 10, 2021
- Source ID
- W81XWH2010797
Entities
People
- Joann Elmore
Organizations
- United States Army
- University of California, Los Angeles