VHU-Net: Variational hadamard U-Net for body MRI bias field correction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Bias field artifacts in MRI cause intensity inhomogeneities, degrading image quality and hindering analysis.
• Accurate bias field correction is crucial for reliable medical image interpretation and downstream tasks.

Purpose of the Study:

• To propose a novel variational Hadamard U-Net (VHU-Net) for effective bias field correction in body MRI.
• To enhance image quality and improve segmentation accuracy through bias field removal.

Main Methods:

• Developed a VHU-Net incorporating convolutional Hadamard transform blocks (ConvHTBlocks) for frequency decomposition and noise suppression.
• Utilized an inverse HT-reconstructed transformer block in the decoder for global, frequency-aware attention.
• Formulated a variational inference-based evidence lower bound (ELBO) for training, promoting latent space sparsity and accurate bias field estimation.

Main Results:

• VHU-Net demonstrated superior performance in intensity uniformity compared to state-of-the-art methods on body MRI datasets.
• Bias field corrected images resulted in substantial improvements in downstream segmentation accuracy.
• The framework showed computational efficiency, interpretability, and robust performance across multi-center datasets.

Conclusions:

• VHU-Net offers an effective and robust solution for MRI bias field correction, suitable for clinical deployment.
• The proposed method enhances image quality and downstream task performance, particularly in segmentation.
• The framework's efficiency and interpretability contribute to its clinical applicability.