Introduction
Epilepsy affects 39 to 50 million people worldwide [
1,
2], about 3–10 per 1000 [
1]. Of these, 30–40% are drug-resistant and need alternative treatments [
2]. Drug-refractory patients with focal epilepsy represent potential candidates to surgical treatment, which consists in the resection of the epileptogenic zone, defined as the site of the beginning of the epileptic seizures. A Cochrane review reported that 65% of about 16,000 patients had a good outcome from surgery [
1], but it strongly depends on accurate localization of the seizure onset zone [
3]. In about 70% of patients, the localization is achieved by combining neuroimaging techniques with noninvasive electrophysiological recordings, such as ElectroEncephaloGraphy (EEG) [
1]. However, EEG does not provide very accurate location of the epileptogenic zone, and especially for drug-resistant epileptic patients, invasive electrophysiological investigations should be carried out. ElectroCorticoGraphy (ECoG) is the most widespread technique to acquire intracranial EEG and is performed by implanting subdural electrodes directly onto the patient's brain surface [
4]. Compared to the EEG electrodes applied on the scalp, the subdural electrodes provide a signal with a much higher resolution and allow a very clear view of the small activity foci [
5]. Subdural electrodes allow not only the localization of abnormal epileptic tissue, but also the localization of adjacent normal functions. Therefore, the precise anatomical localization of the electrodes on the patient’s brain plays a key role in the definition of the epileptogenic zone [
6] or in the mapping of eloquent cortex [
7]. From a clinical point of view, the accurate localization of the anatomical boundaries of the epileptogenic zone allows excluding eloquent areas, avoiding deficits to patient and minimizing brain volume resection. The localization of these electrodes is generally obtained by matching the locations of the electrodes with the brain anatomy of the patient [
8].
Commonly, a pre-implant magnetic resonance image (MRI) is co-registered to a post-implant computed tomography scan (CT) [
9,
10], because MRI offers higher brain tissue contrast, while CT supports electrodes localization [
5], even if CT images are affected by metal artifacts.
Various dedicated software tools that support pre-surgical evaluation are currently available as MATLAB-based packages or open-source software, also with graphical user interfaces. They mainly provide MRI-CT co-registrations and offer only basic features for recognition of ECoG electrodes from CT scans. Synoptic Table
1 reports the most recent software tools, also outlining their main features and limitations for electrodes localization. Most software segments the electrodes via simple image thresholding and requires manual interaction to correct the data. Manual methods are very time-consuming, user-dependent and prone to inaccuracy. On the other hand, the mere CT image thresholding method is not able to recognize all the electrodes and to completely exclude other metal objects, such as wires, tooth fillings, intracranial clips, splinters, stitches, hearing aids and intracranial stents. Hence, manual intervention is often required to adjust the data. The ALICE tool, proposed in [
11], considers the volume of segmented clusters to identify the electrodes, but turned out to be unable to exclude other objects with comparable volumes (e.g., wire clusters).
Table 1
List of software tools for epilepsy pre-surgical evaluation and their approach for electrodes recognition
| Manual + threshold | It is time-consuming and operator-dependent |
| Threshold | Other metallic objects are not excluded |
| Threshold | It requires the neurosurgeon to previously estimate trajectories and target points |
| Threshold | It is not successful in the presence of nearby wires, skull artifacts or overlapping electrodes |
[ 13, 14], iELVis, BioImage Suite | Manual + threshold | It is time-consuming and operator-dependent |
| Manual + threshold | It requires manual selection of electrode voxels and not-electrodes objects must be manually removed |
| Threshold + clustering | It requires manual selection of overlapping or missing electrodes |
| Threshold | It is time-consuming and a semi-manual identification of each electrode centroid must be performed by an expert user |
| Threshold + manual | It requires manual electrodes identification to obtain their 3D coordinates |
This paper presents a novel, more robust, automated method to recognize ECoG electrodes in CT volumes. It consists of identification of metal objects and analysis of their shapes to recognize ECoG electrodes among all detected objects and provide their locations. The proposed approach can be easily implemented in existing tools.
Discussion and conclusions
This study focused on the specific task of automated ECoG electrodes recognition from CT volumes. Currently, only basic automated algorithms are available for this task, which are based on thresholding methods. Indeed, electrodes exhibit higher radiodensities than compact bone, which facilitates their detection by applying a proper threshold on HU values. However, other metal objects (e.g., stiches, clips, connecting wires of electrodes) exhibit such high HU values too, thus impairing electrodes recognition. Moreover, it is well-known that any metal object causes streak metal artifacts on CT images, which may cause alterations of electrodes shapes. Hence, these methods still require intensive, time-consuming manual intervention, to obtain a reasonable accuracy in electrodes localization.
The method presented in this article addresses these issues by means of an automated shape analysis based on machine learning. In particular, the method extracts geometric features of metal objects and applies G-SVM classification to identify the disc-shaped electrodes. It is worth mentioning that the Neuromed database does not include any CT scan acquired with tilted gantry, nor cases with micro-electrodes and penetrating depth electrodes. However, the performances of the proposed method were further assessed on CT volumes from a public repository, which included also penetrating depth electrodes.
The analysis of the results highlighted that the partial volume effect caused an increase in the volume of the electrodes determined from the CT volumes, as compared to their real size. Indeed, the volume of voxel clusters corresponding to an electrode resulted more than double of its real value. This is very evident from the mismatch between the actual electrodes thickness and the length of the tertiary axis. Furthermore, the spatial orientation of electrodes with respect to the CT slice planes caused slight alterations of their shape, which may explain the variability observed in the geometric features. The problem of overlapping electrodes remains unsolved and still requires manual intervention or the adoption of suitable strategies [
11]. In addition, metal artifacts could imply issues in the correct recognition of electrodes with very small inter-distances, since they could be detected as superimposed.
The results obtained both on the 24 Neuromed single-patient datasets and on the seven Mayo single-patient datasets (IDs #12, #16, #20, #22, #26, #28, #31) show a very high percentage of recognized ECoG electrodes with rejection of almost all the other metal objects. The G-SVM average classification accuracies across all patients were 99.74% and 98.27% for the Neuromed and Mayo databases, respectively. The G-SVM achieved comparable performances also on the combined datasets C1 and C2, by scoring classification accuracies of 99.74% and 99.68%, respectively. The higher accuracy accomplished on the combined dataset C2 from the Mayo database, as compared to the related mean accuracy across patients, was reasonably due to the availability of more extensive information on ECoG electrodes features, which led to a more efficient classifier training. Remarkable results were also achieved on the combined datasets by considering a lower number of features, which resulted in classification accuracies in excess of 99%, apart from using the volume alone.
Moreover, the performances of a G-SVM classifier, trained on the combined dataset C1 (Neuromed database), were assessed by testing it on the combined dataset C2 (Mayo database), which had not been used for classifier training. In these tests, the G-SVM scored a mean classification accuracy of 98.94%, which was higher than the one obtained by the tenfold cross-validation performed on the same seven Mayo single-patient datasets. These results confirm, as expected, that the classifier performances benefited from training on a larger dataset, but they also suggest the possibility to apply our method also without training on new data, by using a previously trained classifier. It should be underlined that the limited availability of data for Mayo patients with only ECoG electrodes did not allow to train the classifier on a number of instances comparable to those of Neuromed datasets.
Further tests were carried out on the three additional Mayo single-patient datasets (IDs #5, #17, #27), which also included depth electrodes. A three-class G-SVM classifier was used to separately recognize depth and ECoG electrodes from all other metal objects. The results of the tenfold cross-validation on the combined dataset C3 show a surprisingly high overall percentage of correctly classified objects (~ 99%), with sensitivities to depth and ECoG electrodes, respectively, of 93.75% and 97.93%, and specificities of 99.86% and 99.63%. The three-class G-SVM classifier, previously trained on the combined dataset C3, was tested on the Mayo single-patient dataset #27, which had never been used for classifier training. The classifier scored an overall accuracy of about 96%, with sensitivities to depth and ECoG electrodes, of 83.33% and 97.14%, and specificities of 100.0% and 97.06%. The apparent reduction in the sensitivity to depth electrodes was mainly due to the very small number of instances included in the test set. The encouraging results obtained in depth electrodes recognition by using the same geometric features considered for ECoG electrodes could be ascribed to the CT finite resolution and the partial volume effect, which probably transformed the sleeve-shaped electrodes in full cylinder-like solids. However, the number of depth electrodes the method has been tested on is not as statistically relevant as that of ECoG electrodes in the Neuromed database. Therefore, a more extensive investigation should be carried out in future studies to assess the actual performances of the proposed method for depth electrodes recognition.
The proposed automated method, even when trained with a very limited training set, is able to identify at least the 98% of ECoG electrodes in a CT scan with the 2.7% of misclassified electrodes. When properly trained on a sufficient number of instances, it is able to recognize more than 99% of ECoG electrodes with less than 1% of misclassified electrodes. Hence, to the best of our knowledge, the proposed method achieves unprecedented recognition accuracy, and could provide a substantial reduction in the effort and time consumption required for manual intervention. Moreover, the method proved capable of recognizing depth and ECoG electrodes simultaneously in the same CT volume, thus it could be used also in recent studies that involve both electrodes types. However, in order to attain the highest classification performances, a proper classifier training should be performed which requires the availability of a sufficient number of instances related to electrodes with comparable shape, size and arrangement. Indeed, the use of very few data and/or data obtained with different electrodes can limit the performance of the classifier. Finally, the proposed method can be easily implemented into software suites, such as iELVis [
13] and ALICE [
11], which are widely used to manage the whole preoperative analysis process.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.