Learning from pseudo-labels: deep networks improve consistency in longitudinal brain volume estimation.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Brain atrophy is a key indicator of disease progression in neurodegenerative conditions like multiple sclerosis (MS).
• Accurate measurement of brain atrophy is often hindered by variations in clinical MRI acquisition protocols.
• Developing robust methods is crucial for reliable monitoring of disease and treatment response.

Purpose of the Study:

• To develop and validate a deep learning model, DeepBVC, for accurate brain atrophy measurement in MS.
• To assess the performance of DeepBVC against established methods like SIENA.
• To evaluate the impact of DeepBVC on measuring disease progression in MS patients.

Main Methods:

• A 3D U-Net deep learning model (DeepBVC) was developed using longitudinal 3D T1 MRI scans from MS patients.
• The model was trained to overcome common imaging variances (resolution, SNR, contrast).
• DeepBVC performance was validated against SIENA using test-retest datasets and in-house MS cohorts, correlating with clinical metrics.

Main Results:

• DeepBVC demonstrated superior reproducibility compared to SIENA in test-retest experiments.
• The model showed high consistency across multiple time points in individual MS subjects.
• Both DeepBVC and SIENA showed significant correlation with baseline T2 lesion volume, but not with disability progression.

Conclusions:

• DeepBVC offers a robust, automated, and faster alternative for brain atrophy measurement using T1-weighted MRI.
• Its enhanced performance in handling clinical scan variations makes it suitable for research and clinical use.
• DeepBVC holds potential for improved monitoring of MS progression and treatment effectiveness.