Automatic Delineation of Surgical Resection in Pre- and Post-op MRIs Using a Neural Network
Poster No:
1494
Submission Type:
Abstract Submission
Authors:
Anand Joshi1, Kenneth Taylor2, Takfarinas Medani3, Chinmay Chinara1, Dileep Nair2, Richard Leahy1
Institutions:
1University of Southern California, Los Angeles, CA, 2Cleveland Clnic, Cleveland, OH, 3usc, los angeles, CA
First Author:
Co-Author(s):
Introduction:
Approximately one-third of epilepsy patients are resistant to pharmacotherapy despite the availability of over 20 anti-seizure drugs. If the epileptogenic zone (EZ) is localized, neurosurgery provides a viable therapy to treat these drug-resistant patients, with 40% to 70% becoming seizure-free. The success of the neurosurgery depends on an accurate identification of EZ. Retrospective studies linking presurgical features and resected brain areas can help develop markers that can guide the delineation of surgical resection. These studies require identification of the resection cavity (RC) on preoperative images allowing electrophysiological and imaging features in these regions to be correlated with surgical outcomes. RC segmentation is also crucial in neuro-oncology for estimating the gross tumor volume for radiotherapy.
Methods:
We present a technique for automatically delineating surgical resection on pre-operative MRI using a U-Net neural network. As input, we assume T1-weighted pre- and post-op MRI images. As a preprocessing step, we remove non-brain tissue using our BrainSuite software. The pair of images are then input to a U-Net (Figure 1 a) that performs a non-linear registration of pre- and post-op MRIs. We used masked mean-squared error as the cost function, which allows for the alignment of non-resected brain regions. This is followed by error thresholding and connected component analysis to identify the resection volume in the post-op MRI. The RC is then mapped to the pre-op MRI space using the warping computed by the U-Net.
For validation, we use a subset of the EPISURG dataset [1]. The dataset comprises 430 postoperative MRIs and RCs segmented by three human raters on partially overlapping subsets. Preoperative MRIs are present for 269 subjects, with 53 subjects having both pre- and post-op MRIs. The database also incorporates human-rater segmentations including 33 scans segmented by a researcher in neuroimaging, 34 scans segmented by a clinical research fellow, and 33 subjects segmented by a neurologist. All the scans were also automatically segmented using the proposed method. The automatic and manual delineations were compared using Dice coefficients.
For validation, we use a subset of the EPISURG dataset [1]. The dataset comprises 430 postoperative MRIs and RCs segmented by three human raters on partially overlapping subsets. Preoperative MRIs are present for 269 subjects, with 53 subjects having both pre- and post-op MRIs. The database also incorporates human-rater segmentations including 33 scans segmented by a researcher in neuroimaging, 34 scans segmented by a clinical research fellow, and 33 subjects segmented by a neurologist. All the scans were also automatically segmented using the proposed method. The automatic and manual delineations were compared using Dice coefficients.
Results:
An example resection delineation identified using our technique is shown in Figure 1(b). For the 42 subjects delineated by a researcher in neuroimaging, the average Dice was 0.70. For the subjects delineated by a clinical research fellow, the average Dice was 0.79; for those delineated by a neurologist, the average Dice was 0.74. The histograms of Dice coefficients are shown in Figure 2.
Conclusions:
Based on this initial validation, the higher Dice coefficient for the neurologist and clinical research fellow compared to the neuroimaging researcher is notable, indicating accurate delineations generated by the automated method. Further, use of the warping U-Net allows the RC to be delineated in both pre- and post-op MRIs. The software is documented and integrated into the BrainStorm and BrainSuite software packages. The code is available from https://github.com/ajoshiusc/auto_resection_mask.
Modeling and Analysis Methods:
Image Registration and Computational Anatomy 1
Methods Development
Segmentation and Parcellation 2
Neuroinformatics and Data Sharing:
Workflows
Informatics Other
Keywords:
Epilepsy
Informatics
Machine Learning
MRI
Pre-registration
Segmentation
1|2Indicates the priority used for review