This proposal will help to improve the accuracy of diagnosing melanoma and melanocytic lesions. The incidence of melanoma is rising faster than any other cancer, and ~1 in 50 U.S. adults will be diagnosed with melanoma this year alone. Our research team has noted substantial diagnostic errors in interpreting skin biopsies of melanocytic lesions; pathologists disagree in up to 60% of cases of invasive melanoma, which can lead to substantial patient harm. Our proposal uses computer technology to analyze whole-slide digital images of glass slides in order to improve the diagnosis of melanocytic lesions. Using data from an ongoing NIH study, we will digitize and study a set of 240 skin biopsy cases that includes a full spectrum of benign to invasive melanoma diagnoses. Each biopsy case has a reference consensus diagnosis developed by a panel of three international experts in dermatopathology and new data will be available from 160 practicing U.S. community pathologists, providing a uniquely rich clinical database that is the largest of its kind. This project will include novel computational techniques, including the detection of both cellular-level and architectural entities, the use of a combination of feature-based and deep neural network classifiers, and the use of event graphs with both statistical and structural properties to analyze the accuracy and viewing patterns of pathologists and to determine the characteristics associated with diagnostic errors.
Our specific aims are: 1. To detect cellular-level entities in digitized whole slide images of melanocytic skin lesions. 2. To detect structural (architectural) entities in digitized whole slide images of melanocytic skin lesions. 3. To develop an automated diagnosis system that can classify digitized slide images into one of five possible diagnostic classes: benign; atypical lesions; melanoma in situ; invasive melanoma stage T1a; and invasive melanoma stage ?T1b. 4. To understand physician diagnostic errors and improve their training by investigating image features, pathologist viewing patterns, and the relationship between them, using data from both expert and community pathologists. In our proposed study, we are innovatively merging data on how pathologists review and diagnose slides in a clinical setting using computer image analysis algorithms and machine learning. This technology has the potential to improve the diagnostic accuracy of pathologists by providing an analytical, undeviating review to assist humans in this difficult task. This project will produce tools to reduce these errors, including automated detection and diagnosis tools that suggest regions of interest to pathologists and a visual tracking tool to compare pathologists' search patterns to those of experts, which finds errors in scanning behavior.
Diagnosis of melanoma and melanocytic skin biopsy lesions is among the most challenging areas of pathology and our preliminary data shows concerning levels of errors among pathologists. Our ultimate goal is to use innovative computer image analysis and machine learning techniques to reduce diagnostic errors and save patients' lives. The first step towards this goal is a correct diagnosis of melanoma.
| Elmore, Joann G; Elder, David E; Barnhill, Raymond L et al. (2018) Concordance and Reproducibility of Melanoma Staging According to the 7th vs 8th Edition of the AJCC Cancer Staging Manual. JAMA Netw Open 1: |
