Abstract
Background Accurate surgical resection cavity segmentation on MRI is essential for epilepsy neuroimaging research. Manual segmentation is labour intensive and associated with inter-rater variability. Automated pipelines offer an efficient solution; however, most have been developed for use following temporal epilepsy surgery. We aimed to compare the accuracy of four automated pipelines following either temporal or extratemporal epilepsy surgery.
Methods We identified 4 open-source segmentation pipelines. Epic-CHOP and ResectVol utilise SPM-12 mounted to MATLAB, while Resseg and DeepResection utilise 3D U-net convolutional neural networks. We manually segmented the resection cavity of 50 consecutive patients who underwent epilepsy surgery (30 temporal, 20 extratemporal) using ITK-SNAP. The manual segmentations were compared to the segmentation masks of each automated pipeline using Dice similarity coefficient (DSC).
Results No algorithm correctly identified all resection cavities. ResectVol (n=41, 82%) and Epic-CHOP (n=40, 80%) were able to identify more resection cavities than Resseg (n=22, 44%, P<0.001) and DeepResection (n=21, 42%, P<0.001). Overall, Epic-CHOP had the highest median DSC (0.69), however, this was only statistically significant compared to DeepResection (P<0.01). For the temporal subgroup, Epic-CHOP performed better than DeepResection (P=0.02), but was not different to ResectVol and Resseg. For the extratemporal subgroup, both Epic-CHOP and ResectVol performed better than Resseg and DeepResection (P<0.01), with no difference between Epic-CHOP and ResectVol.
Conclusions Epic-CHOP and ResectVol had high detection rates and accuracy across the cohort. These pipelines could be applied to machine learning studies of outcome prediction to improve efficiency in pre-processing data, however quality control is required.
