Non-invasive diagnosis of melanoma using machine learning and reflectance confocal microscopy: a clinical trial of novel supervised versus weakly supervised algorithms.
File(s)
Author(s)
Kentley, Jonathan
Type
Thesis
Abstract
Melanoma is the third most common skin cancer but accounts for a disproportionate number of deaths. Its incidence continues to rise and early, accurate diagnosis is critical. However, distinguishing melanoma from benign pigmented lesions remains a clinical challenge. Reflectance confocal microscopy (RCM) enhances diagnostic specificity compared with dermoscopy, but its widespread adoption is hindered by a steep learning curve and the time-intensive nature of image acquisition and interpretation.
This research investigates the application of machine learning (ML) to automate both the acquisition and interpretation of RCM images. Three algorithms were evaluated: one for detecting the dermal-epidermal junction (DEJ) to guide RCM stack acquisition, and two for identifying atypical morphological patterns on RCM mosaics, one a fully supervised algorithm and one a weakly supervised learning (WSL) algorithm. These algorithms were tested in three separate studies. Firstly, a retrospective validation study using 141 historical clinical RCM cases (51 melanomas) was performed, followed by a volunteer study to validate an ML-guided imaging pipeline prospectively. Finally, algorithms were evaluated in a prospective patient study including 45 lesions (11 melanomas) imaged as part of clinical care.
Retrospective validation demonstrated high binary diagnostic accuracy for melanoma (area under the receiver operating curve (AUROC) 0.832 for the supervised and 0.803 for the WSL algorithm). In the prospective patient study, the DEJ detection algorithm identified the boundaries of the DEJ with median accuracy of 12μm. The WSL algorithm was statistically non-inferior to the supervised approach (AUROC 0.737 vs. 0.760), supporting its clinical utility despite reduced requirements of annotated data for training.
This work demonstrates the feasibility of a fully automated, ML-guided RCM imaging pipeline. This may allow for wider adoption of RCM and therefore earlier and more accurate diagnosis of melanoma. Further training and development of algorithms using these data is required to allow more widespread adoption.
This research investigates the application of machine learning (ML) to automate both the acquisition and interpretation of RCM images. Three algorithms were evaluated: one for detecting the dermal-epidermal junction (DEJ) to guide RCM stack acquisition, and two for identifying atypical morphological patterns on RCM mosaics, one a fully supervised algorithm and one a weakly supervised learning (WSL) algorithm. These algorithms were tested in three separate studies. Firstly, a retrospective validation study using 141 historical clinical RCM cases (51 melanomas) was performed, followed by a volunteer study to validate an ML-guided imaging pipeline prospectively. Finally, algorithms were evaluated in a prospective patient study including 45 lesions (11 melanomas) imaged as part of clinical care.
Retrospective validation demonstrated high binary diagnostic accuracy for melanoma (area under the receiver operating curve (AUROC) 0.832 for the supervised and 0.803 for the WSL algorithm). In the prospective patient study, the DEJ detection algorithm identified the boundaries of the DEJ with median accuracy of 12μm. The WSL algorithm was statistically non-inferior to the supervised approach (AUROC 0.737 vs. 0.760), supporting its clinical utility despite reduced requirements of annotated data for training.
This work demonstrates the feasibility of a fully automated, ML-guided RCM imaging pipeline. This may allow for wider adoption of RCM and therefore earlier and more accurate diagnosis of melanoma. Further training and development of algorithms using these data is required to allow more widespread adoption.
Version
Open Access
Date Issued
2025-09-22
Date Awarded
2026-06-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Bunker, Christopher
Fearfield, Louise
Sponsor
Melanoma Research Alliance
Grant Number
813399
Publisher Department
School of Public Health
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Medicine (Research) MD (Res)