Disease detection and classification in temporal lobe epilepsy: step-wise versus simultaneous AI decision models in a...

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Diagnostic MRI interpretation for temporal lobe epilepsy (TLE) is subjective.
• Artificial intelligence (AI) offers potential for quantitative support in TLE diagnosis.
• The optimal AI training approach for sequential diagnostic decisions remains unclear.

Purpose of the Study:

• To compare the diagnostic performance of step-wise versus simultaneous AI models for TLE detection and lateralization.
• To evaluate the interpretability and feature focus of different AI model architectures.

Main Methods:

• Analysis of three large epilepsy MRI datasets (n=3676) including people with epilepsy and healthy controls (HC).
• Comparison of step-wise AI models (separate TLE vs HC and L-TLE/R-TLE classification) against a simultaneous model (single classification of HC, L-TLE, R-TLE).
• Utilized EfficientNetV2 model with 3D volumetric T1-weighted images, analyzing class prediction, confidence, and saliency maps.

Main Results:

• Step-wise models significantly outperformed the simultaneous model (P<0.001) in both TLE detection (approx. 2.8% accuracy increase) and lateralization (approx. 12.7% accuracy increase).
• Both models identified similar key features for TLE vs HC discrimination (hippocampus, limbic/cortical regions).
• Step-wise models showed distinct feature focus for lateralization, emphasizing cortical over subcortical regions compared to the simultaneous model.

Conclusions:

• Step-wise AI models demonstrate enhanced diagnostic performance and interpretability for TLE evaluation compared to simultaneous models.
• This approach can improve the early identification of TLE-related structural patterns, aiding timely diagnosis and treatment.
• Future AI clinical support tools can benefit from the step-wise methodology for improved TLE diagnostics.