Robust thalamic nuclei segmentation from T1-weighted MRI using polynomial intensity transformation

Abstract

Accurate segmentation of thalamic nuclei, crucial for understanding their role in healthy cognition and in pathologies, is challenging to achieve on standard T1-weighted (T1w) magnetic resonance imaging (MRI) due to poor image contrast. White-matter-nulled (WMn) MRI sequences improve intrathalamic contrast but are not part of clinical protocols or extant databases. In this study, we introduce histogram-based polynomial synthesis (HIPS), a fast preprocessing transform step that synthesizes WMn-like image contrast from standard T1w MRI using a polynomial approximation for intensity transformation. HIPS was incorporated into THalamus Optimized Multi-Atlas Segmentation (THOMAS) pipeline, a method developed and optimized for WMn MRI. HIPS-THOMAS was compared to a convolutional neural network (CNN)-based segmentation method and THOMAS modified for the use of T1w images (T1w-THOMAS). The robustness and accuracy of the three methods were tested across different image contrasts (MPRAGE, SPGR, and MP2RAGE), scanner manufacturers (PHILIPS, GE, and Siemens), and field strengths (3 T and 7 T). HIPS-transformed images improved intra-thalamic contrast and thalamic boundaries, and HIPS-THOMAS yielded significantly higher mean Dice coefficients and reduced volume errors compared to both the CNN method and T1w-THOMAS. Finally, all three methods were compared using the frequently travelling human phantom MRI dataset for inter- and intra-scanner variability, with HIPS displaying the least inter-scanner variability and performing comparably with T1w-THOMAS for intra-scanner variability. In conclusion, our findings highlight the efficacy and robustness of HIPS in enhancing thalamic nuclei segmentation from standard T1w MRI.

Keywords: Structural imaging; THOMAS; Thalamic nuclei segmentation; Thalamus.

Fig. 1:

The proposed HIPS-THOMAS pipeline. HIPS pre-processing includes normalization of the cropped T1w input, application of a polynomial function to generate a WMn-like image, and a contrast stretching and a rescaling step. The WMn-like cropped input is fed into the THOMAS pipeline as opposed to the original cropped T1w image. Note that for HIPS-THOMAS, the nonlinear warp R uses a cross-correlation metric (CC in blue) and the label fusion step uses joint label fusion (in blue) as opposed to mutual information metric (MI in grey) and majority voting (in grey). NV: normalizing value.

Fig. 2:

HIPS formulation and performance. (a) Normalized intensity plot between WMn-MPRAGE and T1-MPRAGE data from a Philips 3T subject (blue dots) and a 3^rd order polynomial fit (green line). (b) The 10 curves resulting from the individual fitting on 10 Philips data cases (dashed gray) and the resulting aggregated function (red line in a, b). Density plots between normalized WMn-MPRAGE and synthesized WMn-MPRAGE data using the same aggregated function on an example subject from Philips 3T (c), Siemens 3T (d), GE 3T (e), and Siemens 7T (f). The black dashed unity line represents perfect concordance between images. Both axes show normalized voxel intensities.

Fig. 3:

An axial slice from acquired T1w and WMn-MPRAGE as well as HIPS and CNN-synthesized WMn images for a Siemens 3T MPRAGE (a-d), GE 3T SPGR (e-h), Philips 3T MPRAGE (i-l) or Siemens 7T MP2RAGE (m-p) subject with the corresponding nuclei segmentations overlaid on the left thalamus. Note the improved intrathalamic contrast and thalamic boundaries (white arrows on c and g) in the synthesized WMn-like images produced by HIPS-THOMAS compared to the native T1w images. The failure of CNN synthesis can clearly be seen in panels l and p along with the failed or missing (white arrows in l) segmentations. Labels: Table 1

Fig.4:

Box plots of T1w-THOMAS (T1w), HIPS-THOMAS (HIPS), and CNN segmentation’s Dice coefficients (compared against WMn-THOMAS segmentations) for 3T Siemens MPRAGE data (n=12). HIPS-THOMAS significantly improves Dice in whole thalamus and 7 nuclei compared to 1 nucleus for CNN. T-test between the Dice coefficients of the T1w-THOMAS vs. HIPS-THOMAS: P-values *<0.00385 (Bonferonni correction for multiple comparisons, 0.05/13) **<0.001 ***<0.0001. T1w-THOMAS vs. CNN: ¤ p<0.00385 (Bonferonni correction for multiple comparisons, 0.05/13), ¤¤ p<0.001, ¤¤¤ p<0.0001. Labels: cf. Table 1.

Fig.5:

Box plots of T1w-THOMAS (T1w), HIPS-THOMAS (HIPS), and CNN segmentation’s Dice coefficients (compared against WMn-THOMAS segmentations) for GE 3T SPGR data (n=19). HIPS-THOMAS significantly improves Dice in whole thalamus and 7 nuclei compared to 6 nuclei for CNN. T-test between the Dice coefficients of the T1w-THOMAS vs. HIPS-THOMAS: P-values *<0.00385 (Bonferonni correction for multiple comparisons, 0.05/13) **<0.001 ***<0.0001. T1w-THOMAS vs. CNN: ¤ p<0.00385 (Bonferonni correction for multiple comparisons, 0.05/13), ¤¤ p<0.001, ¤¤¤ p<0.0001. Labels: cf. Table 1.

Fig.6:

Box plots of T1w-THOMAS (T1w) and HIPS-THOMAS (HIPS) Dice coefficients (compared against WMn-THOMAS segmentations) for Philips 3T MPRAGE data (n=18). HIPS-THOMAS significantly improves Dice in whole thalamus and 8 nuclei. T-test between the DICE coefficients of the T1w-THOMAS VS HIPS-THOMAS: P-values *<0.00385 (Bonferonni correction for multiple comparisons, 0.05/13) **<0.001 ***<0.0001. Labels: cf. Table 1.

Fig.7:

Box plots of T1w-THOMAS (T1w) and HIPS-THOMAS (HIPS) Dice coefficients (compared against WMn-THOMAS segmentations) for Siemens 7T MP2RAGE data (n=8). HIPS-THOMAS improves Dice in whole thalamus and 9 nuclei. T-test between the Dice coefficients of the T1w-THOMAS VS HIPS-THOMAS: P-values *<0.00385 (Bonferonni correction for multiple comparisons, 0.05/13) **<0.001 ***<0.0001. Labels: cf. Table 1.

Fig.8:

Mean volume error (%) of T1w-THOMAS, HIPS-THOMAS, and CNN segmentations, compared to WMn-THOMAS segmentations for Siemens 3T MPRAGE (a, n=12) and GE 3T SPGR (b, n=19) Philips 3T MPRAGE 3T (c, n=18) and Siemens 7T MP2RAGE (d, n=8) data. The general trend of HIPS < CNN < T1w THOMAS was observed for most nuclei at 3T. HIPS-THOMAS errors were lower than T1w-THOMAS for all nuclei but Hb for Siemens 7T data. Labels: cf. Table 1.

Fig.9:

Graphical summary of improvement of mean Dice coefficients (%) (top row) and reduction of % volume error (bottom row) for each thalamic nucleus from Siemens 3T, GE 3T, PHILIPS 3T and Siemens 7T data using HIPS-THOMAS compared to T1w-THOMAS. Labels: cf. Method 2.2.

Fig.10:

Inter-scanner and intra-scanner variability for the Frequently Traveling Human Phantom MRI dataset (in units of mm3) for T1w-THOMAS, HIPS-THOMAS, and CNN segmentation of the whole thalamus (Thal), AV and VLp nuclei. For this analysis, 24 different scanners covering 3 manufacturers, 2 field strengths and 4 sites each with 3 repeat scans per site were used, resulting in a dataset of 72 scans. Standard deviations (SD) for each method is shown above each plot.