Automatic segmentation of the thalamus using a massively trained 3D convolutional neural network: higher sensitivity for the detection of reduced thalamus volume by improved inter-scanner stability

Abstract

Objectives To develop an automatic method for accurate and robust thalamus segmentation in T1w-MRI for widespread clinical use without the need for strict harmonization of acquisition protocols and/or scanner-specific normal databases. Methods A three-dimensional convolutional neural network (3D-CNN) was trained on 1975 T1w volumes from 170 MRI scanners using thalamus masks generated with FSL-FIRST as ground truth. Accuracy was evaluated with 18 manually labeled expert masks. Intra- and inter-scanner test-retest stability were assessed with 477 T1w volumes of a single healthy subject scanned on 123 MRI scanners. The sensitivity of 3D-CNN-based volume estimates for the detection of thalamus atrophy was tested with 127 multiple sclerosis (MS) patients and a normal database comprising 4872 T1w volumes from 160 scanners. The 3D-CNN was compared with a publicly available 2D-CNN (FastSurfer) and FSL. Results The Dice similarity coefficient of the automatic thalamus segmentation with manual expert delineation was similar for all tested methods (3D-CNN and FastSurfer 0.86 ± 0.02, FSL 0.87 ± 0.02). The standard deviation of the single healthy subject’s thalamus volume estimates was lowest with 3D-CNN for repeat scans on the same MRI scanner (0.08 mL, FastSurfer 0.09 mL, FSL 0.15 mL) and for repeat scans on different scanners (0.28 mL, FastSurfer 0.62 mL, FSL 0.63 mL). The proportion of MS patients with significantly reduced thalamus volume was highest for 3D-CNN (24%, FastSurfer 16%, FSL 11%). Conclusion The novel 3D-CNN allows accurate thalamus segmentation, similar to state-of-the-art methods, with considerably improved robustness with respect to scanner-related variability of image characteristics. This might result in higher sensitivity for the detection of disease-related thalamus atrophy. Key Points • A three-dimensional convolutional neural network was trained for automatic segmentation of the thalamus with a heterogeneous sample of T1w-MRI from 1975 patients scanned on 170 different scanners. • The network provided high accuracy for thalamus segmentation with manual segmentation by experts as ground truth. • Inter-scanner variability of thalamus volume estimates across different MRI scanners was reduced by more than 50%, resulting in increased sensitivity for the detection of thalamus atrophy.

Considerations on brain age predictions from repeatedly sampled data across time

“…T1‐weighted volumes of BBSC1–3 were acquired with a spin echo sequence (TE = 2.95 ms, TR = 6.88 ms, FA = 12°, TI = 450, 188 slices, slice thickness = 1mm, in‐plane resolution = 1 mm × 1 mm, FOV = 256mm, isotropic voxel size = 1 mm 3 ) at a 3T GE system with 32‐channel head coil (see Wang et al., 2022, 2023). T1‐weighted volumes of FTHP1 were acquired at different scanners with various different scanning parameters (see Opfer et al., 2022 or https://www.kaggle.com/datasets/ukeppendorf/frequently-traveling-human-phantom-fthp-dataset). All imaging sites involved in the scanning of FTHP1 were informed that the scan was acquired for the purpose of MRI‐based volumetry.…”

“…We used two datasets for the analyses which had received ethics approval with all participants consenting formally previously (Opfer et al, 2022;Wang et al, 2022Wang et al, , 2023. The first dataset was the Bergen Breakfast Scanning Club (BBSC) dataset (Wang et al, 2022(Wang et al, , 2023, including three male subjects (BBSC2:start-age BBSC2 = 27, BBSC1:start-age BBSC1 = 30, and BBSC3:start-age BBSC3 = 40) who were scanned over the period of circa 1 year with a summer break in the middle of the scanning period (Wang et al, 2022).…”

Considerations on brain age predictions from repeatedly sampled data across time

“…We used two datasets for the analyses which had received ethics approval with all participants consenting formally previously (Opfer et al, 2022;Wang et al, 2022). The first dataset was the Bergen Breakfast Scanning Club (BBSC) dataset (Wang et al, 2022), including three male subjects (BBSC2:start-age BBSC2 = 27, BBSC1:start-age BBSC1 = 30, and BBSC3:start-age BBSC3 = 40) who were scanned over the period of circa one year with a summer break in the middle of the scanning period (Wang et al, 2022).…”

“…This resulted in a total number of N BBSC = 103 scans, relatively equally distributed across subjects (N BBSC1 = 38, N BBSC2 = 40, N BBSC3 = 25). The second dataset was the frequently travelling human phantom (FTHP) MRI dataset (Opfer et al, 2022), including one male subject (FTHP1:start-age FTHP = 48) with 157 imaging sessions at 116 locations, resulting in a total of N FTHP = 557 MRI volumes. Of these, we excluded N = 6 volumes based on errors in the processing pipeline, resulting in a final sample for the main analyses of N FTHP = 551.…”

“…T1-weighted volumes of BBSC1-3 were acquired with a spin echo sequence (TE=2.95ms, TR = 6.88ms, FA = 12°, TI = 450, 188 slices, slice thickness = 1mm, in-plane resolution = 1mm×1mm, FOV = 256mm, isotropic voxel size = 1mm 3 ) at a 3T GE system with 32-channel head coil. T1weighted volumes of FTHP1 were acquired at different scanners with various different scanning parameters (see Opfer et al, 2022 or https://www.kaggle.com/datasets/ukeppendorf/frequentlytraveling-human-phantom-fthp-dataset). All imaging sites involved in the scanning of FTHP1 were informed that the scan was acquired for the purpose of MRI-based volumetry.…”