A pruned and parameter-efficient Xception framework for skin cancer classification.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin cancer is a global health concern, making early detection crucial for patient survival.
• Accurate classification of skin lesions from dermoscopic images is vital for timely diagnosis and treatment.

Purpose of the Study:

• To develop and evaluate a novel framework for enhanced skin cancer classification using dermoscopic images.
• To improve the accuracy and efficiency of automated skin lesion analysis through a combination of advanced machine learning techniques.

Main Methods:

• Utilized transfer learning with models like Xception on the HAM10000 dataset.
• Implemented a layer-based pruning strategy for model optimization and complexity reduction.
• Applied Synthetic Minority Over-sampling Technique (SMOTE) and data augmentation to address class imbalance.
• Employed the Avg-TopK pooling method to preserve critical image features during downsampling.

Main Results:

• Achieved an overall classification accuracy of 91.52%, outperforming several state-of-the-art models.
• Reduced model parameters by approximately 35% post-pruning (from 20.9M to 13.5M), enhancing efficiency.
• Demonstrated significant improvement in model generalization across all skin lesion classes due to SMOTE and data augmentation.

Conclusions:

• The proposed framework effectively combines model pruning, oversampling, and advanced pooling for robust skin cancer classification.
• This approach offers a promising solution for developing efficient and accurate diagnostic tools for clinical application in dermatology.