CAD-PsorNet: deep transfer learning for computer-assisted diagnosis of skin psoriasis

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Image Analysis
• Computational Biology

Background:

• Psoriasis is a chronic, inflammatory skin disorder with diagnostic challenges due to its varied presentation.
• Accurate and timely diagnosis is crucial for effective management of psoriasis.
• The prevalence of psoriasis in northern India's adult population ranges from 0.44% to 2.8%.

Purpose of the Study:

• To develop and evaluate an automated framework for psoriasis detection using deep transfer learning.
• To compare the performance of different deep learning models for psoriasis image recognition.
• To optimize the developed model through hyper-parameter tuning and assess its diagnostic accuracy.

Main Methods:

• A dataset of 325 raw psoriasis images was collected and processed into 496 image patches.
• Four deep transfer learning models (VGG16, VGG19, MobileNetV1, ResNet-50) were utilized for feature extraction and classification.
• Models were adapted with dense, dropout, and output layers; hyper-parameter tuning and AdaGrad optimizer were employed.

Main Results:

• MobileNetV1 initially showed 94.84% sensitivity, 89.37% specificity, and 97.24% accuracy.
• After hyper-parameter tuning, the methodology achieved 94.25% sensitivity, 96.42% specificity, and 99.13% overall accuracy.
• The model's performance, with a Dice coefficient of 0.98, surpassed non-machine learning methods.

Conclusions:

• The developed automated psoriasis image recognition framework demonstrates high accuracy and effectiveness.
• Deep transfer learning, particularly MobileNetV1, shows significant potential for improving psoriasis diagnosis.
• Future work requires diverse datasets to enhance model robustness across different demographics and psoriasis variations.