An Optimized Deep Learning Model for Predicting Mild Cognitive Impairment Using Structural MRI.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Early diagnosis of mild cognitive impairment (MCI) using magnetic resonance imaging (MRI) improves patient outcomes.
• Deep learning models are increasingly used for cost-effective MCI prediction.
• The entorhinal cortex shows early atrophy in MCI, making it a key diagnostic area, though its small size limits research.

Purpose of the Study:

• To develop and optimize deep learning models for differentiating MCI from normal controls using entorhinal cortex MRI data.
• To investigate the efficacy of specific neural network architectures for MCI prediction based on entorhinal cortex features.

Main Methods:

• A dataset focused exclusively on the entorhinal cortex region was constructed for MCI classification.
• Three neural network architectures (VGG16, Inception-V3, ResNet50) were independently optimized for feature extraction.
• A convolution neural network classifier was employed with the optimized architectures.

Main Results:

• The Inception-V3 architecture achieved the best performance for feature extraction.
• Key performance metrics included 70% accuracy, 90% sensitivity, 54% specificity, and 69% area under the curve.
• The model demonstrated an F1 score of 73%, indicating a balanced precision-recall performance.

Conclusions:

• The proposed deep learning approach using entorhinal cortex MRI data is effective for predicting MCI.
• This method shows potential for improving the early diagnosis of MCI, complementing existing diagnostic tools.