Three-dimensional deep learning with spatial erasing for unsupervised anomaly segmentation in brain MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain Magnetic Resonance Images (MRIs) are crucial for diagnosing neurological conditions.
• Unsupervised anomaly detection (UAD) using deep learning offers a way to analyze brain MRIs without pixel-wise annotations.
• Current UAD methods often use 2D slices, neglecting the inherent 3D nature of brain lesions and volumetric spatial context.

Purpose of the Study:

• To investigate if incorporating 3D spatial context via MRI volumes and spatial erasing improves UAD performance over 2D slice-based methods.
• To evaluate and compare 2D and 3D Variational Autoencoders (VAEs) for UAD in brain MRI.
• To assess the impact of dataset size on the performance of 3D UAD methods.

Main Methods:

• Comparison of 2D and 3D Variational Autoencoders (VAEs) for unsupervised anomaly detection.
• Implementation of a novel 3D input erasing technique to enhance spatial context utilization.
• Systematic evaluation of performance based on dataset size.

Main Results:

• 3D VAEs demonstrated superior performance over 2D VAEs, confirming the benefit of volumetric context.
• The proposed 3D erasing methods further improved anomaly segmentation accuracy.
• The best-performing 3D VAE with input erasing achieved a DICE score of 31.40%, outperforming the 2D VAE's 25.76%.

Conclusions:

• 3D deep learning methods, combined with 3D erasing, significantly outperform 2D methods for anomaly segmentation in brain MRI.
• The spatial erasing method enhances performance and reduces the need for extensive datasets.
• This work highlights the advantage of leveraging volumetric data for improved UAD in neuroimaging.