A Novel Morlet Convolutional Neural Network.

Area of Science:

• Medical Technology
• Artificial Intelligence
• Signal Processing

Background:

• Automatic seizure detection is crucial for epilepsy diagnosis and treatment.
• Traditional Convolutional Neural Networks (CNNs) show promise but have limitations like large parameter counts and poor interpretability, hindering edge deployment.
• Existing CNN models struggle with reliability and practical application on resource-constrained devices.

Purpose of the Study:

• To introduce an innovative Morlet convolutional neural network (Morlet-CNN) for effective seizure detection.
• To develop a lightweight and interpretable CNN framework suitable for edge computing.
• To significantly reduce model size and computational requirements while maintaining high accuracy.

Main Methods:

• Developed a Morlet-CNN framework with convolutional kernels having only two learnable parameters for a lightweight architecture.
• Proposed a frequency-domain-response-based kernel pruning algorithm tailored for Morlet-CNN.
• Implemented an INT8 quantization algorithm using Kullback-Leibler (KL) divergence calibration with a Morlet lookup table (LUT).

Main Results:

• Achieved over 90% reduction in model parameter scale through pruning and quantization algorithms with minimal accuracy loss.
• Demonstrated enhanced model interpretability from a signal processing perspective.
• Validated the Morlet-CNN model's efficacy on the Bonn and CHB-MIT datasets, achieving a compact Kilobyte (KB)-level model size.

Conclusions:

• The Morlet-CNN framework offers a highly effective and efficient solution for automatic seizure detection.
• The lightweight and interpretable nature of Morlet-CNN makes it suitable for real-world applications and deployment on edge devices.
• This approach addresses the limitations of traditional CNNs in terms of size and interpretability for epilepsy management.