A Computer Aided Diagnostic Tool for Automatic Psoriasis Segmentation and feature extraction from Dermoscopic images

Background: Psoriasis is a skin disease condition that prevails with red scaly surface along with roughness and requires accurate and precise lesion localization. The subjective methods require more expertize and dermatologist intervention. The time consuming process leads to motivate development of automated tools for detecting psoriasis lesion along with its characteristics which will help in treatment process. Objective: The study presents the deep learning methodology for detection of psoriasis lesion present in dermoscopic images. Methods: A Grad-CAM is applied to highlight discriminative lesion regions from pre-trained classification models. A pseudo ground truth mask is generated with attention maps in order to train U-net segmentation. The attention maps are post-processed to generate pseudo ground truth masks, which are then used to train a U Net segmentation model to detect psoriatic part. The approach eradicates the dependence on pixel-level annotations, aiding scalable lesion mapping. Result: The proposed methodology is applied on dermoscopic psoriasis images obtained from benchmark Dermnet database. The dice coefficient obtained as 95.6. The method accurately localizes and detects the characteristic of psoriasis for further calculate Psoriasis Area Severity Index score. Our approach suggests the potential to help dermatologist as well as patients to detect psoriasis and provide treatment accordingly.