Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Detection Of Acanthosis Nigricans Using Deep Learning: Development And Validation
ABSTRACT
Background:
Acanthosis nigricans (AN) is a skin condition that hints at underlying metabolic derangement. Screening for AN is considered a useful first step in screening for prediabetes and type 2 diabetes. High throughput screening for AN can be benefited by a screening tool that can identify and localize AN regions in images. The You Only Look Once (YOLO) algorithm has revolutionized object detection and can be used for AN screening.
Objective:
The objective of our study was to develop and validate a YOLO-based approach to identify and localize AN in skin images.
Methods:
Development and internal validation of the Acanthosis Detection Assistive Model (ADAM) was conducted using the publicly available dataset offered by Roboflow. The test set from this dataset was further augmented to conduct sensitivity analysis of model performance in response to orientation correction, exposure correction and Gaussian blurring of the images. Subgroup analysis was undertaken to consider the potential influence of skin tone on model performance. Furthermore, the ADAM was externally validated by comparing object localization of the model with that of two experienced clinicians in the field of dermatology.
Results:
The Roboflow dataset provided 560 images for training, 160 for validation and a test set of 80 images for independent internal validation. The ADAM model showed a high accuracy (91% each in the validation and test sets) and F1 score (88% in the validation set and 92% in the test set). For most of the subgroup and sensitivity analyses the model performed consistently well with two exceptions: in the set of images with light skin tones and in the case of images augmented with Gaussian blur, the model had lower accuracy (81% and 88%, respectively). External validation showed that the model localized AN regions comparable to the clinicians’ labels.
Conclusions:
We developed and validated a robust, YOLO-based detection model for AN (the ADAM). Structured screening programs for AN can, conceivably, benefit from this tool. Future studies need to evaluate incorporation of this tool in screening workflow.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.