LV-CadeNet: Long View Feature Convolution-Attention Fusion Encoder-Decoder Network for Clinical MEG Spike Detection

Kuntao Xiao, Xiongfei Wang, Pengfei Teng, Yi Sun, Wanli Yang, Liang Zhang, Hanyang Dong, Guoming Luan, Shurong ShengarXiv 2024 · DOI · arXiv

Abstract

Automated detection of Interictal Epileptiform Discharges (IEDs), or "spikes," is essential for diagnosing epilepsy from long-term EEG and MEG recordings. However, current AI models often suffer from context blindness—failing to consider the temporal surroundings of a waveform—and struggle to capture the complex spatial dipole patterns characteristic of epileptic activity. This paper introduces LV-CadeNet, a novel framework that integrates "long-view" morphological features with a hierarchical convolution-attention encoder-decoder architecture. By normalizing features over a 6-second window and utilizing adaptive feature distillation, LV-CadeNet outperforms significantly larger foundation models. Validated on the TUEV EEG corpus and clinical MEG datasets, the model demonstrates superior accuracy and efficiency, providing a robust tool for clinical epilepsy diagnosis.

Introduction

The diagnosis of epilepsy relies heavily on identifying Interictal Epileptiform Discharges (IEDs) within hours of neurophysiological recordings. Manual review by experts is time-consuming and prone to fatigue. While automated methods exist, they frequently struggle with two primary issues: context blindness and spatial complexity. Human experts typically examine a 5–10 second window to distinguish true spikes from background noise or artifacts. In contrast, many AI models focus on short snippets, missing crucial temporal context. Furthermore, the spatial distribution of potentials across sensors—the "dipole pattern"—is a key diagnostic marker that general-purpose models often fail to model effectively.

Methodology

LV-CadeNet addresses these challenges through a multi-stage pipeline designed to mimic clinical expertise.

The LV-CadeNet pipeline: from raw input data to spatial-temporal feature extraction and final spike classification/detection.

Long-View Feature Construction

A core innovation of LV-CadeNet is the use of biologically-inspired "Long-View" features. Instead of relying solely on raw signal inputs, the model extracts six morphological significance features: topology, amplitude, mean amplitude, slope, half-slope, and sharpness. These features are normalized using a 6-second "long-view" window, providing the necessary temporal context to differentiate pathological spikes from transient artifacts.

CadeNet Architecture

The architecture follows a hierarchical Encoder-Decoder structure:

  • Encoder: Comprises three stages of self-convolution-attention blocks. Convolutions capture local waveform morphology, while attention mechanisms model the spatial relationships across the sensor array. Data is compressed using Pixel Shuffle Downsampling to preserve spatial information.
  • Decoder: Employs cross-convolution-attention for adaptive feature distillation, learning to prioritize diagnostic spatiotemporal patterns over background noise.
Detailed architecture of the LV-CadeNet encoder-decoder stages (a) and the construction of long-view morphological features (b).

Experimental Results

The framework was evaluated on two distinct data types: the public TUEV (Temple University Hospital EEG Events Corpus) for EEG spike classification and proprietary Sanbo-CMR & Sanbo-Clinic MEG datasets for clinical spike detection.

LV-CadeNet demonstrated state-of-the-art performance across both modalities:

  • EEG Classification: On the TUEV dataset, the model achieved a Balanced Accuracy (BAcc) of 67.32%, outperforming the much larger Labram-large model (369M parameters) despite having only 11.79M parameters.
  • MEG Detection: In clinical MEG spike detection, LV-CadeNet reached 86.32% BAcc, a 13.58% improvement over leading baselines.
  • Contextual Robustness: Qualitative analysis confirmed that the "Long-View" module allows the model to correctly reject ambiguous waveforms that appear spike-like in isolation but are clearly noise within a broader temporal context.

Conclusion

LV-CadeNet proves that specialized architectures incorporating domain-specific feature engineering can outperform massive, general-purpose foundation models in neurophysiological tasks. The "Long-View" module is architecture-agnostic and can be integrated into other models to provide immediate performance gains. Future work will focus on improving cross-subject generalizability and assessing performance across diverse patient demographics.

Summary updated

BibTeX
@misc{xiao2024lv,
  author = {Kuntao Xiao and Xiongfei Wang and Pengfei Teng and Yi Sun and Wanli Yang and Liang Zhang and Hanyang Dong and Guoming Luan and Shurong Sheng},
  title = {{LV-CadeNet: Long View Feature Convolution-Attention Fusion Encoder-Decoder Network for Clinical MEG Spike Detection}},
  year = {2024},
  eprint = {2412.08896},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.08896},
}
Figure