Abstract
The skin prick test (SPT) is the primary diagnostic tool for allergies, yet its manual interpretation is often subjective and inconsistent. This study introduces a high-precision automated system for detecting and measuring allergy "wheals" using the Skin Prick Automated Test (SPAT) device. By utilizing a multi-image dataset captured under 32 distinct lighting angles, a deep learning pipeline significantly improves detection accuracy over traditional single-image methods. The proposed approach achieves a Dice coefficient of 0.787 and a detection accuracy of 91.8%, providing a path toward more objective allergy diagnostics.
Introduction
In clinical allergy diagnostics, the skin prick test remains the "gold standard." Clinicians apply allergens to the skin and measure the resulting wheal—a small, raised bump indicating a positive reaction. However, manual measurement is prone to high inter-operator variability, influenced by the clinician's expertise and the tools used. While the Skin Prick Automated Test (SPAT) device was developed to standardize the application of allergens, automated "reading" of the results remains difficult because wheals are often pale and lack contrast against the skin under uniform lighting.
Methods
The researchers developed a two-step automated pipeline to identify and measure wheals from raw image data.
Data Acquisition and Preprocessing
The SPAT device captures a total of 35 images, 32 of which are taken with light sources positioned at different angles around the arm. This setup is designed to create "shadow lines" that highlight the physical topography of the skin.
Pixel-level Classification
The core of the detection system is a modified U-Net architecture. Rather than using complex multi-stream networks to process the 32 images, the authors concatenated them along the channel dimension, resulting in a 96-channel input (32 images x 3 color channels). This allows a standard 2D convolutional network to learn which specific lighting angles and shadow patterns best define the edges of a wheal.
Algorithmic Detection and Matching
To ensure clinical interpretability, the U-Net's output is processed using traditional computer vision techniques. Thresholding and connected components algorithms identify distinct regions. Since the patient's arm may move slightly during the procedure, a rigid transformation (translation and rotation) is applied to align the detected regions with the known theoretical prick locations.
Dataset
The model was trained and validated on a large-scale clinical dataset:
- Participants: 868 adult patients across six hospitals in Belgium and Germany.
- Annotations: 10,416 wheals manually delineated by experts.
- Split: 75% for training (651 patients) and 25% for hold-out validation (217 patients), stratified by hospital.
Key Results
The multi-image approach demonstrated a clear performance advantage over a single-image baseline using uniform lighting:
- Segmentation: The 32-image model reached a Dice coefficient of 0.787, significantly higher than the 0.717 achieved by the baseline.
- Detection: At an Intersection over Union (IoU) threshold of 0.5, the pipeline achieved 91.8% accuracy, compared to 85.2% for the baseline.
- Lighting Sensitivity: Analysis revealed that lighting from the top center and bottom corners of the device provided the most critical information for accurate classification.
Discussion and Limitations
The study highlights that "perfect" uniform lighting is often a disadvantage when detecting raised surfaces like wheals. By treating shadows as valuable data, the system captures topographical information that color and texture alone cannot provide.
However, several limitations remain. The current matching algorithm assumes the arm is a rigid body, whereas skin can stretch and deform. Future work may involve non-rigid registration techniques. Additionally, the method is currently hardware-dependent, requiring the specific lighting array of the SPAT device. Ongoing clinical studies will further evaluate how this automated read-out compares to final diagnoses made by allergists in real-world settings.
Summary updated