Fig 1.
Thalamic Segmentation Workflow.
Overview of the principal stages involved in thalamic segmentation using diffusion-weighted imaging (DWI) and T1-weighted (T1w) images, from preprocessing and feature extraction to clustering and label generation.
Fig 2.
Workflow diagram illustrating clustering.
The thalamus is first divided into small regions using BIRCH clustering for initialization, followed by spectral clustering to produce the final labels.
Fig 3.
Qualitative Analysis of Thalamic Clustering Algorithms.
Axial (z = 6 mm) and coronal (y = –25 mm) sections on the ICBM 2009b T1-weighted template comparing the Krauth-Morel atlas and thalamic parcellation strategy. For each method, the first column shows the spatial probabilistic label maps, and the second column shows the corresponding maximum-probability label maps.
Fig 4.
Evaluation of thalamic segmentation accuracy.
(a) Confusion matrices showing Dice scores between the maximum-probability label maps generated by each clustering method and the reference atlases (Krauth–Morel and Allen Brain Human Atlases). (b) Overlay of maximum-probability label maps and atlas labels on an MNI template, illustrating the anatomical correspondence between data-driven clusters and atlas-defined thalamic nuclei.
Table 1.
Per-nucleus statistical comparison of Dice scores between k-means (nc = 7) and spectral clustering (nc = 9) using Wilcoxon signed ranked test.
Fig 5.
Qualitative Analysis of Pulvinar Parcellation.
Spatial probabilistic label maps and corresponding maximum-probability label maps projected onto the ICBM 2009b T1-weighted MNI template at the axial slice (z = 6 mm) and coronal slice (y = −29). Results are shown for the Krauth–Morel atlas (left), k-means clustering (center), and spectral clustering (right), illustrating qualitative differences in pulvinar parcellation.
Fig 6.
Pulvinar Dice Score Analysis Across Reference Atlases.
Dice score confusion matrices between clustering-derived pulvinar subdivisions and the reference atlases, Krauth–Morel (top) and Allen Human Brain Atlas (bottom). Results are shown for k-means (left) and spectral clustering (right). Rows correspond to atlas labels, and columns correspond to cluster labels.