Deep Learning-Based Enhancement of Already Diagnostic-Quality MRI for Alzheimer's Disease Classification: Effects on...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning (DL) based image enhancement is utilized to improve medical imaging quality.
• Its benefit for diagnostic-quality MRI in downstream tasks and data efficiency is not well understood.

Purpose of the Study:

• Investigate the impact of DL-based enhancement on diagnostic quality structural MRI for Alzheimer's disease (AD) classification.

Main Methods:

• Retrospective analysis of 2293 brain MRI scans from ADNI and 270 from NACC.
• SubtleHD (SHD), an FDA-cleared DL tool, enhanced MRI scans.
• ResNet34 and DenseNet121 models were trained on standard-of-care (SOC) and SHD-enhanced images for AD classification, evaluated by accuracy and macro-AUC.
• Data efficiency assessed by retraining on reduced datasets.

Main Results:

• SHD enhancement significantly improved ResNet34 accuracy (85.2% to 88.7%) and macro-AUC (0.951 to 0.968).
• DenseNet121 showed improved accuracy (90.2% to 92.2%) and macro-AUC (0.978 to 0.982) with SHD enhancement.
• Models trained on 70% of SHD-enhanced data matched full SOC data performance, indicating improved data efficiency.
• In NACC, SHD-trained models significantly outperformed SOC models (accuracy 63.0% vs 49.2%, macro-AUC 0.819 vs 0.679).

Conclusions:

• DL-based enhancement of diagnostic MRI improves Alzheimer's disease classification performance.
• Enhancement reduces the amount of training data needed for machine learning models.
• Conventional image quality metrics may underestimate the information content in enhanced medical images for ML.