Comprehensive brain MRI segmentation in high risk preterm newborns

Abstract

Most extremely preterm newborns exhibit cerebral atrophy/growth disturbances and white matter signal abnormalities on MRI at term-equivalent age. MRI brain volumes could serve as biomarkers for evaluating the effects of neonatal intensive care and predicting neurodevelopmental outcomes. This requires detailed, accurate, and reliable brain MRI segmentation methods. We describe our efforts to develop such methods in high risk newborns using a combination of manual and automated segmentation tools. After intensive efforts to accurately define structural boundaries, two trained raters independently performed manual segmentation of nine subcortical structures using axial T2-weighted MRI scans from 20 randomly selected extremely preterm infants. All scans were re-segmented by both raters to assess reliability. High intra-rater reliability was achieved, as assessed by repeatability and intra-class correlation coefficients (ICC range: 0.97 to 0.99) for all manually segmented regions. Inter-rater reliability was slightly lower (ICC range: 0.93 to 0.99). A semi-automated segmentation approach was developed that combined the parametric strengths of the Hidden Markov Random Field Expectation Maximization algorithm with non-parametric Parzen window classifier resulting in accurate white matter, gray matter, and CSF segmentation. Final manual correction of misclassification errors improved accuracy (similarity index range: 0.87 to 0.89) and facilitated objective quantification of white matter signal abnormalities. The semi-automated and manual methods were seamlessly integrated to generate full brain segmentation within two hours. This comprehensive approach can facilitate the evaluation of large cohorts to rigorously evaluate the utility of regional brain volumes as biomarkers of neonatal care and surrogate endpoints for neurodevelopmental outcomes.

Conflict of interest statement

Competing Interests: The authors have declared that no competing interests exist.

Figures

Figure 1. Manually segmented representative axial T2w slices (# 16–31 of 44 total slices) beginning inferiorly with the amygdalae (cream color), hippocampi (green), brain stem (turquoise), cerebellum (copper) (A–C) and progressing superiorly with thalamic (orange), lenticular (pink), accumbens (blue), and caudate nuclei (lavender) and corpus callosum (yellow) (D–O).

Final 3-dimensional axial and sagittal oblique models of all nine segmented structures (P and Q).

Figure 2. Mid-axial T2w slice on the left highlighting periventricular regions of white matter signal abnormalities (short purple arrows) that the automated segmentation program consistently misclassified as CSF (light blue regions on segmented image on the right).

Occasionally subarachnoid CSF was misclassified as WM (long red arrows).

Figure 3. Commercially available Analyze (white boxes) and in-house developed (IDL environment; gray boxes) segmentation programs were integrated seamlessly to permit various preprocessing and segmentation steps.

The combined manual structural and automated tissue maps were merged in IDL and exported to Analyze for final manual correction and volume calculations.

Figure 4. Mean and standard deviations of automated (left) and semi-automated (right) segmentation accuracy and bias measures, including similarity index (SI), correct estimation index (CEI), over estimation index (OEI), and under estimation index (UEI) for cerebral gray matter (light gray), cerebral white matter (black), and CSF (gray).

Figure 5. A single mid-axial slice exemplifies results at various stages of processing, beginning with unsegmented conventional axial T2 (A) and proton density weighted images (B), automated three tissue segmentation of cerebral GM (gray), cerebral WM (white), and CSF (light blue) (C), manual structural segmentation output (D), combined automated and manual map without correction (E), and final map following manual correction, including relabeling of WM hyperintensities as WMSA (purple) (F).

A 3-dimensional rendering at the same midbrain level displays the relationship of all 13 segmented regions (G).