BMR Detection from Facial Images
Systemic Lupus Erythematosus (SLE) poses significant challenges due to its complex and varied symptoms making diagnosis extremely challenging and time consuming. Symptoms of SLE often mimics other autoimmune or physical conditions and around 5 million people worldwide suffers from this condition, as reported by the Lupus Foundation of American during their study in 2019. However, diagnosis is much more difficult in developing countries with backdated clinical technology and setup therefore, making it virtually unknown the exact number of SLE patient count worldwide. Among all the heterogeneous symptoms presented by SLE, Butterfly Malar Rash is one of the symptoms that is prevalent in the majority of SLE patients, which is a butterfly shaped malar rash (BMR) that appears on the face. This BMR is often misdiagnosed since it is mimicked by other skin conditions like Rosacea, Acne, Eczema, Fifth disease and so on. To speed up the diagnosis process, we plan to design a deep learning models in this project to create a tool for dermatologists that can classify BMR from other similar facial rashes.
Figure: Dataset Generation for Deep Learning Model
Figure: Deep Learning Model for Detection.
Students:
Graduate:
Syeda Lamima Farhat (Graduated Spring 2025)
Shourav Bikash Dey (Graduated Spring 2024)
Undergraduate:
Graham Wilkins
Publications:
[1] A. Jama, G. Wilkins, B. Reamer, N. Haile, N. Mansoor, "Automatic Detection of Butterfly Malar Rash in Facial Images Using Deep Learning", IEEE Conference on Artificial Intelligence, Santa Clara, CA, 2025.
[2] S. Dey, N. Mansoor. "A Butterfly Malar Rash Detection Model for Early Systemic Lupus Erythematosus Diagnosis", 26th International Conference on Computer and Information Technology, 2023. DOI:10.1109/ICCIT60459.2023.10441593
RIPPLES: A Composite Wave Label Embedding Technique for Forward-Forward Algorithm
This project explores a bold new direction in deep learning by extending Geoffrey Hinton’s Forward-Forward algorithm to Convolutional Neural Networks (CNNs) and introducing a novel embedding method called Ripples. Traditional backpropagation has powered many AI breakthroughs, but it comes with major challenges: high computational costs, large memory requirements, and limited biological plausibility. The Forward-Forward algorithm offers a simpler, more biologically inspired approach, opening possibilities for continuous and energy-efficient learning. Ripples is a new composite wave label embedding technique that addresses a critical scalability issue in Forward-Forward research. By generating unique, class-specific patterns, Ripples create richer and more distinct data representations, improving performance on complex image classification tasks like CIFAR-10.
Students:
Graduate:
Spencer Connolly (Expected Graduation in Fall 2025)