BrainLossNet: a fast, accurate and robust method to estimate brain volume loss from longitudinal MRI

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning in Medicine

Background:

• Brain volume loss (BVL) is a key indicator of neurodegeneration.
• Current state-of-the-art methods like SIENA are accurate but computationally intensive.
• There is a need for faster, reliable BVL estimation methods for clinical applications.

Purpose of the Study:

• To introduce BrainLossNet, a novel convolutional neural network (CNN) for rapid and accurate brain volume loss (BVL) estimation.
• To evaluate the performance and robustness of BrainLossNet compared to the established SIENA software.
• To assess the suitability of BrainLossNet for routine clinical neuroimaging workflows.

Main Methods:

• BrainLossNet employs CNN-based non-linear registration for baseline/follow-up 3D-T1w-MRI pairs.
• Brain volume loss (BVL) is calculated using registered brain parenchyma masks, with distortion correction applied.
• The method was trained on 1525 MRI pairs and validated on datasets from multiple sclerosis patients and diverse clinical indications, including phantom data for robustness testing.

Main Results:

• BrainLossNet achieved processing times of 2-3 minutes, significantly faster than SIENA.
• Median brain volume loss (BVL) differences between BrainLossNet and SIENA were minimal (0.08%-0.10%) across different patient cohorts.
• BrainLossNet demonstrated superior robustness to short-term MRI variability compared to SIENA, evidenced by a narrower distribution of apparent BVL in phantom studies.

Conclusions:

• BrainLossNet provides comparable brain volume loss (BVL) estimates to SIENA but with significantly enhanced robustness and speed.
• The method's built-in distortion correction likely contributes to its improved robustness.
• The rapid processing time (2-3 minutes) makes BrainLossNet a viable tool for widespread clinical adoption in routine neurodegeneration assessment.