Objective scoring of psoriasis area and severity index in 2D RGB images using deep learning.

Psoriasis Severity and Area Index (PASI) is a gold standard scoring system for the assessment of Psoriasis skin disease. Generally, PASI scoring is done manually by expert dermatologists through visual and touch senses for psoriasis diagnosis and their treatment's validation. This subjective approach raises several limitations and becomes unreliable. Many conventional and machine learning-based works are proposed for objective estimation of psoriasis area and severity from 2D RGB images. However, these works are validated on small datasets, require manual pre-processing, and rely heavily on hand-crafted features. In the proposed work, a fully automated system based on deep learning is designed for automated PASI scoring from raw 2D RGB images. This system contains a segmentation and three classification models for objective estimation of psoriasis area and severity scores for all three clinical symptoms of psoriasis, respectively. The psoriasis area is estimated by segmenting healthy and unhealthy regions simultaneously using a lightweight network as a backbone with UNet. After segmentation, the severity scores for each segmented lesion are automatically estimated by using a hybrid classification model. This model is developed by adopting a lightweight network for local feature extraction and integrating it with a vision transformer for learning global features. The psoriasis dataset used in the proposed work is self-prepared and contains 1,018 photographic images from different body regions of 212 psoriasis patients. The exhaustive performance analysis is done for the automatic estimation of each parameter of PASI. The proposed work achieves mean absolute error of 0.04, 0.23, 0.22, and 0.21 for objective estimation of Area, Redness, Scaliness, and Thickness scores, respectively. The mean absolute error obtained by the proposed system for automatic scoring of PASI is 1.02. The comparative studies with existing works further validate the efficacy of the proposed work. This work can further be improvised by using data from multi-centre and regions in a large population. [ABSTRACT FROM AUTHOR]