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

TL;DR

DeepBVC is a deep learning method that improves the accuracy and consistency of longitudinal brain volume change measurements in MRI scans, outperforming traditional methods like SIENA especially across different scanners and acquisition protocols.

Contribution

This paper introduces DeepBVC, a novel deep learning approach trained with pseudo-labels to enhance robustness and consistency in brain volume change estimation from MRI scans.

Findings

DeepBVC outperforms SIENA in consistency across multiple time points.

DeepBVC is robust to variations in imaging parameters.

DeepBVC offers faster and automated measurements.

Abstract

Brain atrophy is an important biomarker for monitoring neurodegeneration and disease progression in conditions such as multiple sclerosis (MS). An accurate and robust quantitative measurement of brain volume change is paramount for translational research and clinical applications. This paper presents a deep learning based method, DeepBVC, for longitudinal brain volume change measurement using 3D T1-weighted MRI scans. Trained with the intermediate outputs from SIENA, DeepBVC is designed to take into account the variance caused by different scanners and acquisition protocols. Compared with SIENA, DeepBVC demonstrates higher consistency in terms of volume change estimation across multiple time points in MS subjects; and greater stability and superior performance in scan-rescan experiments. Moreover, the results also show that DeepBVC is insensitive to acquisition variance in terms of imaging contrast, voxel resolution, random bias field and signal-to-noise ratio. Measurement robustness, automation and processing speed suggest a broad potential of DeepBVC in both research and clinical quantitative neuroimaging applications.