Evaluating Deep Learning for Cerebellar Peduncle Segmentation: Assessing TractSeg’s Performance
Poster No:
247
Submission Type:
Late-Breaking Abstract Submission
Authors:
Soumen Ghosh1, Susmita Saha2, Thiago Rezende3, Ian Harding4
Institutions:
1QIMR Berghofer, Brisbane, QLD, 2Monash University, Melbourne, Victoria, 3Department of Neurology, University of Campinas, Campinas, Brazil, 4QIMR Berghofer Medical Research Institute, Brisbane, Queensland
First Author:
Co-Author(s):
Introduction:
The cerebellar peduncles (CPs), comprising the inferior cerebellar peduncle (ICP), middle cerebellar peduncle (MCP), and superior cerebellar peduncle (SCP), are essential white matter tracts facilitating crucial bidirectional communication between the cerebellum and other brain regions. These pathways play vital roles in both motor and non-motor functions, and their involvement in rare neurodegenerative conditions, such as Friedreich Ataxia (FRDA) [1], spinocerebellar ataxia [1] and multiple system atrophy [2], underscores the importance of accurate CP segmentation. Studies using diffusion tensor imaging for tract reconstruction [3] struggle with complex regions like the SCP decussation, where crossing fibers cause segmentation errors [4]. Previous studies [4] highlight two challenges: time-consuming manual segmentation with inter-rater variability and disease-specific models limiting generalizability.
TractSeg, a fully convolutional neural network, has shown promise in efficient white matter segmentation using fiber orientation distribution function peaks [5]. Deep learning models like TractSeg enable fast segmentation, but their robustness in complex tracts like CPs needs validation, especially in cerebellar neurodegeneration. This study critically evaluates the accuracy and reliability of TractSeg for CP segmentation against manually segmented ground truth data in healthy and FRDA cohorts.
TractSeg, a fully convolutional neural network, has shown promise in efficient white matter segmentation using fiber orientation distribution function peaks [5]. Deep learning models like TractSeg enable fast segmentation, but their robustness in complex tracts like CPs needs validation, especially in cerebellar neurodegeneration. This study critically evaluates the accuracy and reliability of TractSeg for CP segmentation against manually segmented ground truth data in healthy and FRDA cohorts.
Methods:
Our dataset included high-resolution T1-weighted and diffusion MRI scans from FRDA patients (n=3) and healthy controls (n=3), acquired on a 3T MRI scanner across three sites. Diffusion MRI data were acquired using a multiband EPI sequence (voxel size: 1.5 mm³, TR = 3230 ms, TE = 89 ms, MB = 4). The protocol included 197 diffusion directions across two shells (b = 1500, 3000 s/mm²) and 6 b0 images. Acquisition time is 5'37'' (AP) + 5'37'' (PA). Diffusion-weighted Imaging (DWI) data underwent a preprocessing pipeline utilizing QSIPrep, correcting for motion, eddy current, and susceptibility distortions. We used manually delineated ground truth masks of the ICP, MCP, and SCP in native diffusion (B0) space, achieved using Principal Eigenvector maps [6], and the Linear Westin Index map [7]. A domain expert meticulously created the masks using these maps and the T1-weighted images co-registered to B0 space. For quality control, a second evaluator randomly selected six CPs from six subjects, yielding a correlation of 0.83 ± 0.05 (95% CI: [0.77, 0.88]). The pre-trained TractSeg model was used for automated CP segmentation, and its performance was evaluated using the Dice Similarity Coefficient (DSC) against manual segmentations as the gold standard.
Results:
Quantitative analysis revealed sub-optimal performance of TractSeg in CP segmentation compared to manually generated CPs, with moderate performance in ICP segments but less consistency in MCP and SCP. Visual inspection (Figure 1) of the TractSeg output indicates difficulties in accurately segmenting the cross-section of the SCP, particularly in regions with complex fiber orientations and crossing fibers.
Table 1 presents a quantitative comparison of TractSeg's segmentation performance for CPs across different subjects, grouped by study site and diagnosis (FRDA or control). TractSeg CP segmentations have marginal overlap with the ground truth, with DSCs ranging from 0.16 to 0.40, indicating rough localization. Our analysis suggests that, despite its efficiency, TractSeg may not generate accurate CP maps, particularly in FRDA patients (mean DSC are lower except for MCP), due to anatomical variations and geometric complexity.
Table 1 presents a quantitative comparison of TractSeg's segmentation performance for CPs across different subjects, grouped by study site and diagnosis (FRDA or control). TractSeg CP segmentations have marginal overlap with the ground truth, with DSCs ranging from 0.16 to 0.40, indicating rough localization. Our analysis suggests that, despite its efficiency, TractSeg may not generate accurate CP maps, particularly in FRDA patients (mean DSC are lower except for MCP), due to anatomical variations and geometric complexity.
·Visual interpretation of decussation-SCP. The first column shows the manually segmented ground truth, while the second column displays the region of interest generated by the pretrained TractSeg model
·Comparison of TractSeg's performance in cerebellar peduncles segmentation against manually segmented ground truth.
Conclusions:
These findings suggest that a "one-size-fits-all" deep learning approach such as TractSeg might not be ideal for the accurate segmentation of complex structures such as the cerebellar peduncles or for all disease cohorts. This study highlights the need for tailored training datasets, like tractography-based manual segmentation, to enhance ground truth for deep learning model training.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 2
Methods Development
Segmentation and Parcellation
Keywords:
Cerebellum
Computational Neuroscience
Machine Learning
MRI
Neurological
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Nicoletti, G., Fera, F., Condino, F., Auteri, W., Gallo, O., Pugliese, P., ... & Quattrone, A. (2006). MR imaging of middle cerebellar peduncle width: differentiation of multiple system atrophy from Parkinson disease. Radiology, 239(3), 825-830.
3. Vanderah, T. W., & Gould, D. J. (2016). Nolte's The human brain: an introduction to its functional anatomy (pp. 126-153). Philadelphia, PA, USA:: Elsevier.
4. Ye, C., Yang, Z., Ying, S. H., & Prince, J. L. (2015). Segmentation of the cerebellar peduncles using a random forest classifier and a multi-object geometric deformable model: application to spinocerebellar ataxia type 6. Neuroinformatics, 13, 367-381.
5. Wasserthal, J., Neher, P., & Maier-Hein, K. H. (2018). TractSeg-Fast and accurate white matter tract segmentation. Neurolmage, 183, 239-253.
6. Fan, X., Thompson, M., Bogovic, J. A., Bazin, P. L., & Prince, J. L. (2011, August). A novel contrast for DTI visualization for thalamus delineation. In Proceedings of SPIE--the International Society for Optical Engineering (Vol. 7625, p. 762533).
7. Ying, S. H., Landman, B. A., Chowdhury, S., Sinofsky, A. H., Gambini, A., Mori, S., ... & Prince, J. L. (2009). Orthogonal diffusion-weighted MRI measures distinguish region-specific degeneration in cerebellar ataxia subtypes. Journal of neurology, 256(11), 1939-1942.