Poster R15 - Comparison of GRACE and SynthSeg Deep Learning Models for Whole-Head Brain Segmentation

Introduction: Transcranial direct current stimulation (tDCS) shows promise as a treatment for neurological diseases; however, improvements must be made to improve consistency. One reason for variability is anatomical differences amongst patients, such as variations in skull thickness and brain shape. Whole-head brain segmentation of individuals can be used to plan targeted treatments that optimize the tDCS parameters. Manual or semi-automatic whole-head brain segmentation is work-intensive and time-inefficient, taking 22-30 hours per individual head. Deep learning models can provide fully automated segmentation tasks to counteract this. This work trained two deep learning segmentation pipelines from the GRACE [1] and SynthSeg [2] models using Magnetic Resonance Imaging (MRI) scans of older adults to evaluate their performances.

Materials and

Methods: This investigation employed a dataset derived from the Augmenting Cognitive Training in Older Adults (ACT) trial, which tested the effectiveness of cognitive training in conjunction with tDCS for cognitive improvement [3]. The subset from this work comprised 177 MRI images, each with dimensions of 176x256x256 voxels. All participants were cognitively healthy older adult subjects aged between 65 and 89. Each image was segmented into 11 tissue labels: white matter, grey matter, eyes, cerebrospinal fluid (CSF), air, major artery (blood), cancellous bone, cortical bone, skin, fat, and muscle. This process was performed by trained manual segmentors using a semi-automatic process. All images were extensively reviewed for quality. The dataset was partitioned into 157 training volumes and 20 testing volumes. Initial experimentation involved training both the SynthSeg and GRACE models on the training set using the recommended parameters from each model’s paper [1-2], followed by the utilization of the trained model to segment each image within the testing set. Subsequently, the predicted segmentations were evaluated against the reference annotations using established metrics such as the DICE score and Hausdorff distance.

Results, Conclusions, and Discussions: Our analysis reveals that the GRACE model outperforms SynthSeg in terms of alignment with ground truth annotations, as evidenced by a lower average Hausdorff distance and higher average DICE score. Lower Hausdorff distances and higher DICE scores indicate a more accurate model. GRACE appears to be the best at the tissues that are optimal for tDCS simulation. Nonetheless, it is noteworthy that SynthSeg exhibits comparatively lower variability, as indicated by the absence of outlier observations in our analysis. However, it is essential to acknowledge the inherent dissimilarities in model architectures. SynthSeg operates as a self-supervised model trained on unlabeled data, whereas GRACE benefits from training on labeled data. Thus, as our data shows, SynthSeg is more likely to underperform on the in-distribution testing set, but it may perform better on unseen images. Inference time for GRACE is about 3 seconds per image, while SynthSeg takes about 10 minutes. Comparative analysis between the GRACE model and SynthSeg demonstrates superior performance of the former. However, it is imperative to contextualize these findings within the framework of differing model architectures to attain a more nuanced understanding of the observed disparities. The use of deep learning models for segmentation can allow for more accurate and rapid parameter estimation in tDCS intervention.

Acknowledgements (Optional): [1] S. E. Stolte et al., “Precise and rapid whole-head segmentation from magnetic resonance images of older adults using Deep Learning,” Imaging neuroscience (Cambridge, Mass.), https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10922731/
[2] B. Billot et al., “Synthseg: Segmentation of brain MRI scans of any contrast and resolution without retraining,” ScienceDirect, https://www.sciencedirect.com/science/article/pii/S1361841523000506
[3] A. Woods, R. Cohen, M. Marsiske, and G. Alexander, “Augmenting cognitive training in older adults (the ACT study): Design and methods of a phase III tdcs and cognitive training trial,” Contemporary Clinical Trials, https://www.sciencedirect.com/science/article/pii/S1551714417305219?via%3Dihub.