Psoriasis severity classification based on adaptive multi-scale features for multi-severity disease.

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Psoriasis evaluation relies on subjective clinician assessment using the Psoriasis Area and Severity Index (PASI).
• Existing methods struggle with the variable and irregular severity patterns characteristic of psoriasis, especially in multiple-severity cases.

Purpose of the Study:

• To develop a novel, objective method for evaluating psoriasis severity.
• To improve the accuracy of psoriasis assessment, particularly for complex, multi-severity presentations.

Main Methods:

• Generated multi-severity psoriasis images using CutMix data augmentation.
• Proposed a hierarchical multi-scale deformable attention module (MS-DAM) for adaptive region detection.
• Integrated MS-DAM with EfficientNet B1 for psoriasis classification.

Main Results:

• The EfficientNet B1 model with MS-DAM achieved a high F1-score of 0.93.
• MS-DAM demonstrated over 5% higher accuracy compared to the multi-scale channel attention module (MS-CAM).
• Gradient-weighted activation mapping confirmed the method's alignment with human visual perception.

Conclusions:

• The proposed MS-DAM offers a more objective, effective, and accurate approach to psoriasis severity analysis.
• This AI-driven method enhances the evaluation of diverse psoriasis presentations, including those with irregular severity.
• The findings suggest a significant advancement over traditional subjective assessment methods.