Soft Sensing of Local Skin Reactions: A Computer Vision Approach Using Deep Model Colour Space Analysis for Objective...

Allergic contact dermatitis (ACD) is a prevalent condition traditionally diagnosed using standardized patch testing, a method limited by subjective visual skin evaluation and delayed skin reaction manifestations. This study introduces a novel "soft optical sensor" methodology that automates the localization and quantification of local skin reactions. The proposed computer vision framework employs a hybrid deep learning ensemble, integrating DeepLabV3+, UperNet, and SegFormer architectures for the precise segmentation of healthy skin and the detection of patch test markers. To objectively evaluate the intensity of patch-tested reactions, the system transforms extracted regions of interest (ROIs) into the CIELAB colour space. A personalized healthy skin baseline is computed for each patient, allowing the measurement of absolute colour deviation to classify skin reactions into three distinct severity categories (small, medium, and large). Evaluated on a clinical dataset, the hybrid ensemble demonstrated high segmentation performance, achieving a peak Dice score of 89.3% and an intersection over union (IoU) of 81.3%. Furthermore, the colour space analysis showed high congruence with a single expert dermatologist's assessment (90% overall accuracy), particularly in unambiguous reaction extremes. This automated approach offers a scalable, objective, and standardized solution for clinical dermatological evaluations.