A Lightweight Breast Cancer Mass Classification Model Utilizing Simplified Swarm Optimization and Knowledge...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Breast cancer is a growing global health concern, necessitating effective early detection methods.
• Current deep learning models for breast cancer classification are often computationally intensive, limiting their accessibility.
• There is a need for efficient and lightweight models that perform well under resource constraints.

Purpose of the Study:

• To develop an optimized, lightweight deep learning model for breast mass abnormality classification.
• To address the limitations of large-scale, computationally expensive models in breast cancer detection.
• To improve the cost-effectiveness and accessibility of AI-driven breast cancer diagnostic tools.

Main Methods:

• Utilized the CBIS-DDSM dataset for training and validation.
• Developed a novel concatenated classification architecture with a two-stage strategy.
• Employed data augmentation, image preprocessing, knowledge distillation, and Simplified Swarm Optimization (SSO).

Main Results:

• The proposed lightweight model outperformed standalone Convolutional Neural Network (CNN) and Deep Neural Network (DNN) models.
• Knowledge distillation significantly improved the compact model's performance.
• The final SSO-Concatenated NASNetMobile (SSO-CNNM) model achieved a 96.17% compression rate and high performance metrics (96.47% accuracy, 97.4% precision, 94.94% recall, 98.23% AUC).

Conclusions:

• The developed lightweight model offers a computationally efficient and highly accurate solution for breast mass abnormality classification.
• The two-stage strategy combining knowledge distillation and SSO effectively optimizes deep learning models for resource-constrained environments.
• This research provides a promising approach for enhancing early breast cancer detection through accessible AI technologies.