D2.77 - A Novel Neural Network for Quantifying Wheal and Erythema in Skin Prick Testing

Quantification of allergic tests is crucial for evidence-based allergology. Recent advances in deep neural network technologies have introduced innovative methods for various medical applications. However, applying these techniques to skin prick testing in allergology presents two key challenges: (1) the limited availability of annotated training datasets, which often results in overfitting and biased outcomes; and (2) interference from preexisting skin conditions—such as dermatitis, urticarial rash, pigmented spots and tattoos—that can result in false-positive signals.

We propose a novel pipeline that integrates paired pre- and post-allergen exposure images, automatically registers these images, and analyzes them using a Kolmogorov-Arnold Network (KAN). This approach aims to reliably detect wheals and erythema while addressing the challenges of small training datasets and spurious signals from existing skin abnormalities. Our methodology begins with high-resolution imaging of the test area under standardized conditions. Pre-test and post-test images are automatically registered to align anatomical features, enabling the neural network to isolate changes specifically induced by the allergenic challenge. Despite being trained on a limited dataset of only a few hundred instead of usual tens-to-hundreds of thousands images, the KAN-based architecture has been optimized to mitigate overfitting and bias. Additionally, by analyzing the differences between registered image pairs, the neural network effectively suppresses false-positive signals arising from preexisting skin abnormalities.

Our experimental data demonstrated that the KAN-based model does not show signs of overfitting and exhibits high accuracy in detecting wheals and erythema. This neural network–based approach accelerates and streamlines the quantification process, thereby reducing reliance on subjective manual evaluations.

The integration of automatic registration of skin images with the KAN presents an efficient solution for quantifying allergic reactions in skin prick testing. Proposed pipeline overcomes the challenges of limited annotated datasets and false-positive signals from preexisting skin abnormalities, supporting more accurate evidence-based diagnostics in allergology. Future work will focus on implementing this pipeline in routine clinical diagnostics and exploring its potential for therapy control.