Methodological Approaches to Machine Learning Based on a Small Sample Size for Contributing Assessment to the Decision Support System in Dermatology

In the field of dermatology, the diagnostic process is often difficult due to a wide variety of skin conditions requiring the expertise of a dermatologist, who are currently in very insufficient numbers. Since 2003, tele dermatology initiatives have been implemented in Madagascar, but local experience has shown unsatisfactory results due to the unavailability of specialists to respond to requests posted on the platforms. Over the past five years, the literature is full of several studies proposing Artificial Intelligence (AI) algorithms to aid in diagnosis. Thus, our study aims to seek a better approach to diagnostic support by conducting a meta-analysis of experiences combining telemedicine and artificial intelligence in dermatology and to propose a model for a diagnostic support system adapted to our context. The first draft of model training results with U-Net for 458 annotated images presented in this article shows training parameters better suited to a small sample of images. The result gave us a stable accuracy on average at 97.83% from the 5th epoch of training. With the data collected as part of the passion dermatology project, which contains 1960 photos of skin lesions, our solution focuses on (a) a semantic segmentation model proposing a tool for evaluating knowledge about skin lesions and (b) a diagnostic orientation mechanism. The model thus obtained will become a self-assessment tool for students and physicians to visually familiarize themselves with skin pathologies in the absence of a dermatologist. The pre-trained machine will thus become a medical training tool to improve the diagnostic approach.

Author(s): Andriambololoniaina Faly Herizo1,2,3*, Randriamanalina Rojo Nofidiantsoa1, Gottfrois Philippe5, Rakotoarisaona Mendrika Fifaliana2, Fenohasina Georges Cheny2, Ramily Leophonte2, Rakotomanana Kiady Armando2, Ratovohery Andry Nirilalaina1,3, Randriamboavonjy Rado Lalao1,3, Bagayoko Cheick Oumar4, Rakotomalala Mamy Alain1, Rapelanoro Rabenja Fahafahantsoa2,3, Hsu Christophe5, Navarrini Alexander5, Ravonimanantsoa Ndaohialy Manda-Vy1

Abstract | Full-Text | PDF

Share This Article