Deep Learning-Based High-Frequency Ultrasound Skin Image Classification with Multicriteria Model Evaluation

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• High-frequency ultrasound (HFUS) imaging offers novel insights into skin structures and pathologies.
• Accurate classification of skin conditions like atopic dermatitis and psoriasis is crucial for effective treatment.
• Deep learning models show promise for analyzing complex medical images.

Purpose of the Study:

• To apply convolutional neural networks (CNNs) for the first time to HFUS skin image classification.
• To develop a robust framework for segmenting skin layers and classifying pathologies.
• To introduce a classification confidence measure for enhanced reliability of deep learning models in dermatology.

Main Methods:

• A database of 631 HFUS skin images (healthy and pathological) was curated.
• Epidermal layer segmentation was performed using DeepLab v3+ with an Xception backbone.
• Transfer learning was utilized for segmentation and feature extraction.
• Five classification models were trained and evaluated using different data augmentation strategies.
• A novel confidence measure combining skin layer maps and Grad-CAM heatmaps was introduced.

Main Results:

• The DenseNet-201 model, utilizing the segmented region of interest, demonstrated superior classification accuracy and reliability.
• The proposed classification confidence level effectively indicated the reliability of the deep learning model's predictions.
• A multicriteria evaluation measure facilitated the selection of the optimal model based on accuracy, confidence, and dataset size.

Conclusions:

• CNNs are effective tools for classifying HFUS skin images, supporting dermatological diagnosis.
• The developed framework, including segmentation and confidence estimation, enhances the reliability of AI in skin imaging.
• The DenseNet-201 model with region of interest extraction represents a promising approach for automated skin pathology classification.