Unique Presentation Identifier:

39

Program Type

Graduate

Faculty Advisor

Dr. Robin Ghosh

Document Type

Poster

Location

Face-to-face

Start Date

9-4-2026 1:00 PM

End Date

9-4-2026 3:00 PM

Abstract

According to the World Health Organization's press release on December 12, 2024, global healthcare spending is dropping significantly, leaving a large percentage of the world without proper healthcare. In an attempt to alleviate this problem, with respect to the field of dermatology, we created a deep learning model, Dermatology Enhanced by Recognition and Machine Aided Learning (DERMAL), to assist in diagnosing skin conditions. DERMAL was trained on a portion of the Google and Stanford Medicine's SCIN dataset, which has more than 10,000 images of various skin conditions. The 9 most common skin conditions of the dataset were selected as the training portion. The diagnosis labels for each record were made by 3 volunteer dermatologists who each gave their possible diagnoses. These 3 separate diagnosis were then aggregated and converted to percentages to form the skin condition labels. To generate a more balanced dataset, new images were generated from the originals by applying random augmentation to minority classes. Those images and a list of their symptoms were trained on by a neural network. The model structure has a CNN branch to train on the input images, and an MLP branch to train on the given symptomatology, which are then concatenated to provide a multi label output. The best validation accuracy obtained was 0.8856. The future of this project involves the creation of a web application; this next step will allow DERMAL to provide accessible, low-cost preliminary screening tools to undeserved communities.

Comments

Keywords: dermatology, deep learning, medical image classification, multi- input neural network, symptom-based modeling, healthcare accessibility, skin condition detection, computer vision, AI-assisted diagnosis

Download

Share

COinS
 
Apr 9th, 1:00 PM Apr 9th, 3:00 PM

DERMAL: A Multi-Input Deep Learning Model for Improving Access to Dermatological Screening

Face-to-face

According to the World Health Organization's press release on December 12, 2024, global healthcare spending is dropping significantly, leaving a large percentage of the world without proper healthcare. In an attempt to alleviate this problem, with respect to the field of dermatology, we created a deep learning model, Dermatology Enhanced by Recognition and Machine Aided Learning (DERMAL), to assist in diagnosing skin conditions. DERMAL was trained on a portion of the Google and Stanford Medicine's SCIN dataset, which has more than 10,000 images of various skin conditions. The 9 most common skin conditions of the dataset were selected as the training portion. The diagnosis labels for each record were made by 3 volunteer dermatologists who each gave their possible diagnoses. These 3 separate diagnosis were then aggregated and converted to percentages to form the skin condition labels. To generate a more balanced dataset, new images were generated from the originals by applying random augmentation to minority classes. Those images and a list of their symptoms were trained on by a neural network. The model structure has a CNN branch to train on the input images, and an MLP branch to train on the given symptomatology, which are then concatenated to provide a multi label output. The best validation accuracy obtained was 0.8856. The future of this project involves the creation of a web application; this next step will allow DERMAL to provide accessible, low-cost preliminary screening tools to undeserved communities.