Deep Learning-Driven Transformation: A Novel Approach for Mitigating Batch Effects in Diffusion MRI Beyond...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Magnetic Resonance (MR) image "batch effects" from varying scanner hardware and parameters compromise image quality.
• These variations hinder the generalizability of deep learning (DL) models used in medical image analysis.

Purpose of the Study:

• To develop a DL model for contrast adjustment and super-resolution to standardize diffusion-weighted images (DWIs).
• The goal is to reduce diversity in DWIs caused by different magnetic field strengths and imaging parameters.

Main Methods:

• A DL model was trained and validated on a dataset of 1134 adult subjects using data from seven MR scanners (1.5T and 3T).
• The model employed contrast adjustment and super-resolution techniques to harmonize DWI data.
• Evaluation involved radiologist assessment, image quality metrics (PSNR, SSIM), texture analysis, and a ResNet-50 model performance comparison.

Main Results:

• The DL protocol successfully reduced variations in DWI contrast and resolution across different MR devices.
• Performance metrics for a ResNet-50 model showed a significant decrease in accuracy, precision, recall, and F1 score after harmonization, indicating reduced machine-specific bias.
• t-SNE visualization confirmed improved feature consistency across scanners, and the autoencoder halved learning iterations with a Dice coefficient >0.74 for lesion signal reproducibility.

Conclusions:

• The developed DL strategy effectively mitigates batch effects in diffusion MR images.
• This approach improves the quality and generalizability of MR images for DL applications in radiology.