Self-supervised Diffusion-based 3D Brain MRI Enhancement for Anomaly Detection and Segmentation

Abstract

Magnetic resonance imaging (MRI) is a widely used non-invasive imaging technique for brain image acquisition in medical diagnostic analysis and treatment planning. During acquisition, MRI data are corrupted by various factors, such as random Gaussian, Rician, and Rayleigh noise probabilities, signal distortions, and motion artifacts inherent to the MRI system. Denoising is a major pre-processing imaging task in computer-aided diagnostic (CAD) systems applied to suppress the noise effect and regenerate a high-quality enhanced image with maximum useful image information. Traditional filtering-based MRI denoising and restoration methods demonstrate limited-quality image generation with single-image denoising techniques, whereas these methods fail to preserve essential information, such as edge features, in the restored images with large input datasets. Deep-learning-based (DL) image denoising methods have demonstrated significant performance in natural image restoration applications. DL methods have great potential to preserve and enhance the image quality of reconstructed images without losing important image information. The increasing complexity of medical image analysis techniques underscores the need for robust and advanced anomaly detection and automatic medical image feature classification. Existing methods have certain limitations and face challenges in identifying and capturing the desired information patterns of anomalies and often limit their usage to specific types of lesions in brain MRI. This study addresses the challenge of using a DL-based novel unsupervised learning ‘reverse-encoder–decoder network model to generate different types of MRI pathologies. The proposed unsupervised anomaly detection methods were evaluated using BraTS2021, IXI, and simulated BrainWeb MRI datasets and demonstrated improved anomaly detection, image restoration accuracy and segmentation performance in terms of low error and higher peak signal-to-noise ratio (PSNR), structural similarity (SSIM), dice score metrices values.