Knowledge Distillation-Based TinyML Model for Breast Cancer Detection Using Real and Wasserstein GAN-Generated...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Tiny Machine Learning (TinyML) enables intelligent cancer diagnostics on edge devices.
• Optimizing TinyML for breast cancer diagnosis using microwave imaging data presents domain-specific challenges.
• Scarcity of microwave imaging datasets hinders the development of robust diagnostic models.

Purpose of the Study:

• To propose a novel Knowledge Distillation framework for efficient TinyML models in breast cancer diagnosis.
• To enhance the accuracy and reduce the model size of TinyML diagnostic tools.
• To address the data scarcity issue by generating a synthetic microwave imaging dataset.

Main Methods:

• A teacher-student model approach using residual convolutional neural networks for knowledge distillation.
• Development of a lightweight student model trained via knowledge transfer from a high-accuracy teacher model.
• Generation of a synthetic breast cancer dataset using Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and validation with a one-class Support Vector Machine.

Main Results:

• The teacher model achieved 95.42% accuracy.
• The student model, after knowledge distillation, reached 95.32% accuracy with a 96% reduction in model size.
• The student model without distillation achieved 86.5% accuracy, highlighting the effectiveness of the proposed method.
• Training on combined real and synthetic data enhanced model robustness.

Conclusions:

• The proposed Knowledge Distillation framework significantly improves the efficiency and accuracy of TinyML models for breast cancer detection.
• The synthetic dataset generation addresses data scarcity, enhancing the viability of TinyML for real-world breast cancer diagnostics.
• This approach facilitates efficient and accurate breast cancer detection on edge devices, supporting early diagnosis in clinical settings.