Deep Learning Application for Effective Classification of Different Types of Psoriasis.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Psoriasis is a prevalent chronic inflammatory skin condition affecting 125 million people globally.
• Deep learning shows promise in diagnosing and classifying skin diseases, including various psoriasis types.

Purpose of the Study:

• To develop and evaluate a deep learning application for classifying five types of psoriasis (plaque, guttate, inverse, pustular, erythrodermic) and normal skin.
• To compare the performance of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) algorithms for psoriasis classification.

Main Methods:

• Utilized 172 normal skin images (BFL NTU dataset) and 301 psoriasis images (Dermnet dataset).
• Applied image preprocessing techniques: data augmentation, enhancement, and segmentation.
• Extracted features including color, texture, and shape.
• Trained CNN and LSTM models on 80% of the dataset.

Main Results:

• The CNN model achieved an accuracy of 84.2%.
• The LSTM model achieved an accuracy of 72.3%.
• A paired sample T-test showed a statistically significant difference (p < 0.001) between CNN and LSTM accuracies.

Conclusions:

• The developed deep learning application demonstrates significant potential for accurate psoriasis classification.
• CNN outperformed LSTM in classifying psoriasis types, suggesting its suitability for dermatological applications.
• This approach may be adaptable for broader dermatological prediction tasks.