Authors :
Presenting Author: Yejin Ann, BA – Center for Neuroscience Imaging Research, Institute for Basic Science, Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University, Suwon, South Korea
Seok-Jun Hong, PhD – Center for Neuroscience Imaging Research, Institute for Basic Science, Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University, Suwon, South Korea / Center for the Developing Brain, Child Mind Institute, NY, United States; Youn-Min Sohn, MD, PhD – Neuroscience Center, Samsung Medical Center, Seoul, South Korea / Department of Neurology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, South Korea
Rationale:
The anterior and centromedian thalamic nuclei (ATN, CM) are the commonly target regions in deep brain stimulation (DBS) for epilepsy. To optimize this surgical planning, several MRI-based thalamic atlases have been introduced, but they were mostly unsuccessful for the purpose of precise DBS targeting (Ewert et al., 2018). Moreover, while the efficacy of DBS has been reported in many previous studies, most were based on a small thalamic region-dependent seizure reduction (SR), not a rigorous imaging-based prediction, which may guarantee generalizable accuracy for unseen cases. Here, we addressed these issues by 1) validating existing thalamic atlases based on histological comparisons, 2) building a convolutional neural network (CNN) predicting SR based on DBS patterns, and 3) assessment of the overlap of significant regions (from this prediction) with the previously validated atlases.
Methods:
We validated existing thalamic atlases (Allen (Ding et al., 2017), Freesurfer (Iglesia et al., 2018), Ilinsky (Ilinsky et al., 2018), Distal (Ewert et al., 2018) and Thomas (Su et al., 2019)) by comparing the postmortem brain data (Alkemade et al., 2022). We first manually segmented ATN and CM in histology based on their anatomical landmarks (Fig1A) and calculated a dice index (DI) between these segmentations and the borders of nuclei in MRI atlases. We then chose the atlas with the highest DI in further analyses.
Next, we built a CNN-based classifier to take the responder apart from the non-responder group (RG/NRG) for new patients based on the DBS patterns. We analyzed 50 patients including 31 focal (ATN) and 19 generalized epilepsy (CM). To avoid an arbitrary SR threshold to determine RG/NRG, we varied the threshold between 30-70%, splitting the patients into the two groups. After five fold cross validations, we selected the threshold providing the highest accuracy and generated an ‘attention’ map, using a gradCAM (Gotkowski et al., 2021), to examine the importance of subregions for prediction. Finally, we computed DI between attention maps and atlases.
Results:
Our validation demonstrated that ATN and CM in histology better match to different atlases. Indeed, the ATN in histology showed the largest overlap with the Thomas, while in CM, the Freesurfer showed the best-fit boundary (Fig1B). In the CNN experiment, the model exhibited the highest accuracy for RG/NRG prediction at 30% threshold in ATN and at 70% in CM (Fig 2A). Thus, we generated the attention map with these thresholds and compared them with the selected atlases (Fig 2B). The ATN showed the moderate DI (0.33) for RG, which is nevertheless >2 times larger than those in NRG (0.16). Conversely, the CM showed similar DI for both groups (RG; 0.31, NRG; 0.34; Fig2C).
Conclusions:
Our study validated the existing MRI-based thalamic atlases based on postmortem histology. According to our CNN-based analyses, the DBS targeting ATN seems moderately efficacious. However, in CM DBS it is hard to tell the difference between RG/ NRG by relying on the nucleus level. Therefore, subsequent studies looking at the whole brain connectivity beyond the thalamic nuclei should be warranted.
Funding:
Institute for Basic Science