Deep Learning-based Segmentation of Laser Thermal Therapy Ablation Volume
Abstract number :
2.331
Submission category :
9. Surgery / 9C. All Ages
Year :
2023
Submission ID :
569
Source :
www.aesnet.org
Presentation date :
12/3/2023 12:00:00 AM
Published date :
Authors :
Presenting Author: Christian Raimondo, MS – Sidnety Kimmel Medical College at Thomas Jefferson University
Mahdi Alizadeh, PhD – Department of Neurosurgery – Thomas Jefferson University; Raphael Fernandes Casseb, PhD – University of Campinas; Brunno de Campos, PhD – University of Campinas; Bahram Hejrani, MD – Department of Neurosurgery – Thomas Jefferson University; KiChang Kang, BS – Department of Neurosurgery – Thomas Jefferson University; Fernando Cendes, MD, PhD – University of Campinas; Lara Jehi, MD – Cleveland Clinic; Chengyuan Wu, MD, MSBmE – Department of Neurosurgery – Thomas Jefferson University
Rationale: Laser interstitial thermal therapy (LiTT) is a minimally invasive therapeutic intervention for mesial temporal lobe epilepsy (mTLE) which has been shown to be a viable alternative to anterior temporal lobectomy (ATL). The location of the LiTT ablation cavity and the specific structures involved have been shown to be important in maximizing procedural efficacy and risk reduction. Current methods for determining post-operative ablation volume require manual segmentation of resected areas, which are time-intensive and include variation between clinicians. Deep learning-based semantic segmentation methods have demonstrated robust performance in discriminating target regions across multi-modal imaging datasets. ResectVol is a validated automated segmentation tool for identifying surgically resected regions in a cohort of mTLE patients. This study applies a beta upgraded version of ResectVol to a cohort of LiTT mTLE patients and evaluates performance by comparing the automated segmentation region to manual segmentations from two independent neurosurgeons.
Methods: A nnU-NetV1 LiTT model (nnU-Net) was trained in an independent Cleveland Clinic and UNICAMP cohort of patients who underwent open surgical resection, expanding from the published ResectVol segmentation tool developed previously. Automated segmentation using this model was performed on post-operative T1-weighted MRI scans from 10 patients who underwent LiTT for mTLE treatment at Thomas Jefferson University. Manual segmentation was performed by two experienced neurosurgeons (Grader 1, Grader 2). Masks generated from manual segmentation were treated as ground-truths, independently. Segmentation performance was evaluated using Dice-Sorensen coefficient (DSC) and Jaccard similarity to assess the overlap between nnU-Net and the two assessors. A Wilcoxon signed-ranked test was used to determine differences between automated and manual segmentations. Volumes between the groups were assessed using One-way ANOVA. For quality assurance, all three masks were visually inspected by overlaying on the T1-weighed image in ITK-snap before analysis [Fig 1].
Results: There was no statistical difference in average DSC or Jaccard scores between automated nnU-Net segmentation vs Grader 1 (mean DSC = 0.76; mean Jaccard = 0.65) and automated nnU-Net segmentation vs Grader 2 (mean DSC = 0.75; 0.625671, p = 0.32) [Fig 2]. The average volume segmented by the nnU-Net mask was smaller (13,204 cm³) compared to manual segmentation methods (Grader 1 = 17,327 cm³; Grader 2 = 19,002 cm³), but this difference was not statistically significant (p = 0.29) [Fig 2].
Conclusions: Our findings indicate automated segmentation of ablation volume following LiTT procedure using nn-Unet based tool is comparable to manual segmentation. These results validate the application of this model for tasks involving LiTT and demonstrate the benefits of deploying deep-learning tools for epilepsy surgery research.
Funding: The authors declare that no funding was received for this study.
Surgery