Adding the third dimension: 3D convolutional neural network diagnosis of temporal lobe epilepsy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Convolutional neural networks (CNNs) show promise for analyzing structural abnormalities in temporal lobe epilepsy (TLE).
• Three-dimensional (3D) CNNs often surpass two-dimensional (2D) CNNs in medical image analysis.

Purpose of the Study:

• To investigate if 3D CNNs outperform 2D CNNs in identifying TLE-specific MRI features.
• To analyze the impact of sample size and clinical characteristics on model performance.

Main Methods:

• Trained and compared 3D and 2D CNNs on 1178 T1-weighted MRI scans (589 TLE, 589 controls).
• Utilized feature visualization to identify critical regions for classification.
• Evaluated model performance across varying sample sizes and with image harmonization.

Main Results:

• The 3D CNN achieved a median accuracy of 86.4%, significantly outperforming the 2D CNN (median 83.0%) in 84% of comparisons.
• The 3D CNN's advantage was evident only with large sample sizes.
• Both models highlighted medial-ventral temporal, cerebellar, and midline subcortical regions, with 3D CNN showing higher salience.

Conclusions:

• 3D CNNs are valuable for detecting subtle MRI abnormalities in TLE, particularly in patients lacking visible lesions like hippocampal sclerosis.
• The superior performance of 3D CNNs is contingent upon training with substantial datasets.