Deep learning-based automated segmentation of resection cavities on postsurgical epilepsy MRI

Abstract

Accurate segmentation of surgical resection sites is critical for clinical assessments and neuroimaging research applications, including resection extent determination, predictive modeling of surgery outcome, and masking image processing near resection sites. In this study, an automated resection cavity segmentation algorithm is developed for analyzing postoperative MRI of epilepsy patients and deployed in an easy-to-use graphical user interface (GUI) that estimates remnant brain volumes, including postsurgical hippocampal remnant tissue. This retrospective study included postoperative T1-weighted MRI from 62 temporal lobe epilepsy (TLE) patients who underwent resective surgery. The resection site was manually segmented and reviewed by a neuroradiologist (JMS). A majority vote ensemble algorithm was used to segment surgical resections, using 3 U-Net convolutional neural networks trained on axial, coronal, and sagittal slices, respectively. The algorithm was trained using 5-fold cross validation, with data partitioned into training (N = 27) testing (N = 9), and validation (N = 9) sets, and evaluated on a separate held-out test set (N = 17). Algorithm performance was assessed using Dice-Sørensen coefficient (DSC), Hausdorff distance, and volume estimates. Additionally, we deploy a fully-automated, GUI-based pipeline that compares resection segmentations with preoperative imaging and reports estimates of resected brain structures. The cross-validation and held-out test median DSCs were 0.84 ± 0.08 and 0.74 ± 0.22 (median ± interquartile range) respectively, which approach inter-rater reliability between radiologists (0.84-0.86) as reported in the literature. Median 95 % Hausdorff distances were 3.6 mm and 4.0 mm respectively, indicating high segmentation boundary confidence. Automated and manual resection volume estimates were highly correlated for both cross-validation (r = 0.94, p < 0.0001) and held-out test subjects (r = 0.87, p < 0.0001). Automated and manual segmentations overlapped in all 62 subjects, indicating a low false negative rate. In control subjects (N = 40), the classifier segmented no voxels (N = 33), <50 voxels (N = 5), or a small volumes<0.5 cm³ (N = 2), indicating a low false positive rate that can be controlled via thresholding. There was strong agreement between postoperative hippocampal remnant volumes determined using automated and manual resection segmentations (r = 0.90, p < 0.0001, mean absolute error = 6.3 %), indicating that automated resection segmentations can permit quantification of postoperative brain volumes after epilepsy surgery. Applications include quantification of postoperative remnant brain volumes, correction of deformable registration, and localization of removed brain regions for network modeling.

Keywords: Automated segmentation; Convolutional neural network; Hippocampal remnant; Postoperative MRI; Resection cavity; Temporal lobe epilepsy.

Fig. 1

Schematic of evaluation metrics of classifier performance. Three metrics were applied: Dice-Sørensen coefficient (DSC), Hausdorff distance (HD), and manual versus automated segmented volumes (VOL). DSC quantifies the overlap between manual and resected segmentations in a range of 0 to 1. HD quantifies the farthest distance between the boundary points. The number of voxels and voxel size quantifies the volume. The classifier performance was optimized by maximizing the DSC.

Fig. 2

Pipelines for automated resection segmentation and quantification of postsurgical volume estimates. The resection segmentation pipeline uses a U-Net architecture (top) and produces a 3D binary mask of resected tissue. To quantify postoperative remnant volumes (bottom), the preoperative image was segmented into brain regions. The intersection of the resection and anatomical brain segmentations were used to generate a resection report.

Fig. 3

Classifier accuracy across the cross-validation cohort. Here we report model performance on the held-out test sets (N = 45) during cross-validation. (A) Dice-Sørensen coefficient (DSC), 0.84 ± 0.08 (median ± interquartile range). (B) 95 % Hausdorff distance, 3.61 ± 2.64 mm (median ± interquartile range). (C) Pearson correlation between predicted and manually segmented volumes (r = 0.94, p < 0.0001). (D) Representative manual and automated segmentations from each quartile of the Dice score distribution. Segmentations are overlaid on the T1-weighted images, with their associated DSC on the right-hand side.

Fig. 4

Classifier accuracy across the held-out cohort. Here we report model performance on the final held-out test set (N = 17) collected after model development. (A) Dice-Sørensen coefficient (DSC), 0.74 ± 0.22 (median ± interquartile range). (B) 95 % Hausdorff distance, 4.04 ± 10.32 mm (median ± interquartile range). (C) Pearson correlation between predicted and manually segmented volumes (r = 0.87, p < 0.0001). (D) Representative manual and automated segmentations from each quartile of the Dice score distribution. Segmentations are overlaid on the T1-weighted images, with their associated DSC on the right-hand side.

Fig. 5

Example output from the lowest scoring segmentations. (A) The lowest segmentation overlap case in the cross-validation set was a subject with hyperintense blood product in the resection cavity. (B) In the held-out test set, the lowest segmentation overlap was a SAH case, where surgical tracts were manually segmented but not included in the automated segmentation.

Fig. 6

Graphical User Interface (GUI) for estimating surgical remnants. Here we illustrate the GUI interface developed for estimating resection remnants on a selective amygdalohippocampectomy patient. (A) In the first panel, the user selects to run the full pipeline or run the analysis using a resection volume they generated. (B) The user then uploads the required images and selects their desired registration and segmentation parameters. (C) The pipeline outputs a table of affected regions by percentage resected and provides an interactive visualization of the resection segmentation for manual review and quality control.

Fig. 7

Strong correlation between remnant estimates using automated and manual methods. We compared hippocampal remnant estimates using automated and manual resection segmentations. Automated and manual estimates are significantly correlated (r = 0.90, p < 0.0001) and have a mean absolute error of 6.3 %.