Assist-Dermo: A Lightweight Separable Vision Transformer Model for Multiclass Skin Lesion Classification.

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Image Analysis
• Computer Vision for Healthcare

Background:

• Automated classification of pigmented skin lesions (PSLs) aids early skin cancer detection.
• Existing systems often require high computational resources, limiting deployment on constrained devices.
• There is a need for accurate, yet computationally efficient, deep learning models for PSL classification.

Purpose of the Study:

• To develop an automatic classification system, Assist-Dermo, for recognizing nine classes of PSLs.
• To design a separable vision transformer (SVT) architecture with fewer parameters and comparable accuracy to state-of-the-art (SOTA) models.
• To improve runtime performance and diagnostic efficacy for clinical experts.

Main Methods:

• Developed a novel SVT architecture integrating SqueezeNet and depthwise separable convolutional neural network (CNN) models.
• Employed data augmentation to address PSL imbalance and pre-processing for lesion region selection and enhancement.
• Utilized a diverse dataset including Ph2, ISBI-2017, HAM10000, and ISIC for training and evaluation.

Main Results:

• Achieved high performance metrics: 95.6% accuracy (ACC), 96.7% sensitivity (SE), 95% specificity (SP), and 0.95 area under the curve (AUC).
• The Assist-Dermo system demonstrated superior performance compared to SOTA algorithms in classifying nine PSL classes.
• The model's efficiency was validated through its reduced parameter count and improved runtime performance.

Conclusions:

• The Assist-Dermo system offers a computationally efficient and accurate solution for classifying pigmented skin lesions.
• It effectively assists dermatologists in early skin cancer detection through dermoscopy.
• The developed model code is publicly available on GitHub, promoting further research and development.