JParc: Joint cortical surface parcellation with registration

TL;DR

JParc is presented, a joint cortical registration and parcellation framework, that outperforms existing state-of-the-art parcellation methods and can significantly increase the statistical power in brain mapping studies as well as support applications in surgical planning and many other downstream neuroscientific and clinical tasks.

Abstract

Cortical surface parcellation is a fundamental task in both basic neuroscience research and clinical applications, enabling more accurate mapping of brain regions. Model-based and learning-based approaches for automated parcellation alleviate the need for manual labeling. Despite the advancement in parcellation performance, learning-based methods shift away from registration and atlas propagation without exploring the reason for the improvement compared to traditional methods. In this study, we present JParc, a joint cortical registration and parcellation framework, that outperforms existing state-of-the-art parcellation methods. In rigorous experiments, we demonstrate that the enhanced performance of JParc is primarily attributable to accurate cortical registration and a learned parcellation atlas. By leveraging a shallow subnetwork to fine-tune the propagated atlas labels, JParc achieves a Dice score greater than 90% on the Mindboggle dataset, using only basic geometric features (sulcal depth, curvature) that describe cortical folding patterns. The superior accuracy of JParc can significantly increase the statistical power in brain mapping studies as well as support applications in surgical planning and many other downstream neuroscientific and clinical tasks.

Figures (7)

• Figure 1: Graphical representation of the probabilistic model. Circles are random variables. Rounded squares indicate parameters. Shaded (gray) quantities are observations. The big plate represents replication over subjects ($i$). The subscript $p$ stands for parcellation and the subscript $g$ stands for geometry. $\bm{A}_{p}$ represents the parcellation atlas, $\tilde{\bm{A}_{p}}$ the deformed parcellation atlas (in the subject space) by the parcellation-specific deformation field $\bm{\phi}_{p}$. The operator $(\;\cdot\;)^{-1}$ indicates a warp inverse -- a warp from the atlas space to the subject space. $\bm{I}_{g}$ is the observed subject geometric image. $\bm{I}_{p}$ is the observed subject parcellation map. The yellow shaded area indicates the registration and $\bm{\theta}_{p}$ is the parcellation prediction function (fine-tune).
• Figure 2: Overview of JParc network architecture. The JOSA registration module (yellow box) takes individual geometric features $\bm{I}_{g}$ and propagates the parcellation atlas $\bm{A}_{p}$ to the individual space through the parcellation-specific deformation. The deformed parcellation atlas $\tilde{\bm{A}}_{p}$, concatenated with the individual geometric features $\bm{I}_{g}$, are input to a parcellation head $\bm{\theta}_{p}$ to generate the final prediction on the parcellation $\hat{\bm{I}}_{p}$. The predicted parcellation $\hat{\bm{I}}_{p}$ is compared to the manual parcellation $\bm{I}_{p}$ for loss evaluation.
• Figure 3: Visual comparison of parcellation maps for an exemplar subject. Lateral surface view is shown on top and medial surface view is shown on the bottom for each method. Yellow arrows indicate regions where JParc substantially outperforms other baseline methods in comparison to the manual parcellation map.
• Figure 4: Bar plots of Dice coefficients for each of the 32 ROIs. The ROIs' names are shown along the x-axis and the Dice score is shown on the y-axis. The box represents the median Dice and the black vertical lines through the center of each box indicate the interquartile range (middle 50%). Baseline methods are shown in various blue shades and JParc is shown in orange.
• Figure 5: Examples of unsatisfactory parcellation for some ROIs. Cases for parsorbitalis, precentral, and anterior cingulate are shown in the left, middle, and right columns, respectively. The top row shows the JParc result and the bottom row shows the manual parcellation map. The Dice for the corresponding ROI and the overall ROI for that subject are shown below the images in each column.