Deep Learning-based Assessment of Internal Carotid Artery Anatomy to Predict Difficult Intracranial Access in Endovascular Recanalization of Acute Ischemic Stroke

Of the 447 patients included initially, 68 had to be excluded due to unavailability of CTA. Among the remaining 379 patients, 31 patients had a tandem occlusion (concomitant occlusion of the proximal or cervical ICA). Data of these patients were used for the development of the DL models as the vascular anatomy was still valid for learning purposes; however, in the analysis of procedural times in relation to the vessel angle, patients with tandem occlusion or inadequate contrast filling of the ICA, either due to higher grade proximal stenoses that were not treated in the acute setting, or due to late contrast phases, were not included. Figure 1 shows an overview of the patient selection.

The CTA data were exported from the picture archiving and communicating system (PACS) to a local machine for further analysis. Table 1 summarizes the distribution of the Y‑axis voxel spacings which were in the majority 0.5 mm in the Y‑axis voxel spacing (n = 221). CTA data of a lower Y‑axis resolution was also considered to obtain networks that perform on heterogeneous input regarding the imaging quality.

For the development of the DL models a train/test split was performed, with 15% of the cases being allocated to the test set resulting in 321 train and 58 test cases.

In order to achieve information reduction and to specifically select anatomical structures for classification, a DL-based vessel segmentation was developed, 121 CTAs were segmented manually (3D Slicer with a 3-dimensional sphere brush [26]) and 3 distinct labels were assigned to the different anatomic sectors during manual segmentation. Label 1 was given to the aorta, label 2 to the left CCA + ICA, and label 3 to the right CCA + ICA.

Of the labeled cases, 100 belonged to the training set and 21 to the test set. On the training set, a nnUNet [27] was trained. The training mode used was 3d_lowres with the nnUNetTrainerV2_noMirroring trainer class. The training ran for 1000 epochs. The resulting model was evaluated on the test set using Dice score, precision, and recall.

In order to make use of the trained nnUNet for the segmentation of the remaining 258 CTAs, a hybrid segmentation approach was conceived. This approach consisted of 2 steps:

Manual verification was performed by overlaying the predicted segmentation onto the respective CTA. The integrity of the segmentations was checked in this fashion, and corrections were performed if necessary.

In order to obtain a training set as large as possible for the DL development, both ICAs of each patient were used. Left and right ICAs were assessed regarding the most acute angle in their course. The assessment was performed using the angle measurement tool in 3D Slicer. The measurement was performed using the angle measurement tool in 3D Slicer which enables angle measurement on the surface of segmentations. The standard CTA reconstruction planes often do not match the plane of the vessel course. By setting the angle measurement points on the vessel surface a plane is projected onto it. This plane is a close approximation of the true plane of the course of the vessel. It therefore allows for more precise angle measurement than on standard CTA reconstruction planes.

Figure 2 shows the measurement of two exemplary angles.

The continuous angles were explored regarding their distribution as well as correlation with procedural times. Furthermore, a cut-off analysis to find optimal thresholds for classification was performed. For this purpose, we divided all cases in “long” cases with groin puncture to first maneuver above the 80th percentile (41 min and longer) as well as groin puncture to final TICI above the 80th percentile (113 min and longer). Optimal thresholds of ICA angles for best discrimination between “long” cases versus all other cases were calculated using receiver operating characteristic curves by maximizing Youden’s index and running 1000 bootstrap repeats. The performance of the thresholds was assessed using a machine-learning classification algorithm. On basis of these analyses, ICAs were grouped into two classes (acute angle ≤ 90° or obtuse angle > 90°). This dichotomization was performed to enable classification with a neural network as well as to ensure comparability of results with previous works [18]. Of the ICAs 66 were not assessable as they were not continuously contrasted in the CTA (either due to late contrast phases or proximal higher grade stenoses that were not treated in the acute setting). After class formation, the association between the manually measured ICA angle on the occluded side in relation to EVT duration, as well as the interval from groin puncture to first thrombectomy maneuver was explored.

To facilitate the following DL-based classification task, a preprocessing was designed, consisting of the following steps:

A 3D CNN was designed according to Zunair et al. and Ertl et al. [28, 29]. Figure 3 gives an overview of the general network setup. A Rectified Linear Unit was used as the activation function except for the last layer, which has a sigmoid activation function. Binary Cross Entropy was used as the loss function and Adam as the optimizer. The training was performed over 100 epochs with a learning rate of 0.0001 and a batch size of 2. In terms of data augmentation, rotation in two planes was applied stochastically and with random angles and mirroring. The input for the network was the segmentation of a CCA + ICA. The output of the network was 0 or 1 with 0 relating to the class ICA angle 90° and 1 to the class ICA angle ≤ 90°. The network was implemented with keras and tensorflow (DL frameworks for the python programming language).

The training set specified after data export was further split into five folds for cross-validation. Five models were trained, each time using a different fold as the validation set and the other four as the training set. Training was performed on a machine running Ubuntu 20.04 (Canonical, London, United Kingdom) with 6 cores, 64 GB RAM and an Nvidia (Nvidia, Santa Clara, CA, USA) Titan RTX.

Each of the five trained models predicted on the test set and their predictions were averaged. Performance was assessed by accuracy and AUC.

The ICA angle predictions of the model on the test set were then again explored regarding association in relation to EVT duration as well as the interval from groin puncture to first maneuver.

Evaluation against the test set of the trained nnUNet resulted in Dice scores of 0.944, 0.854, and 0.856 for the aorta, left CCA/ICA, and right CCA/ICA, respectively, summarized in Table 3. Table 4 shows the time efforts for manual and hybrid segmentation. Prediction of a segmentation took 30 s with another 3 min on average for manual verification. The hybrid segmentation was 11 times faster than manual segmentation. Figure 4 gives an overview of the segmentations resulting from the trained nnUNet.

Figure 5 shows the distribution of ICA angles on the occluded side and Table 5 gives an overview about descriptive measures regarding the manually measured angles on the occluded side. The correlation analysis between ICA angles and procedural time showed a significant but weakly negative correlation for EVT duration (Pearson’s r = −0.12, p < 0.022), while for the correlation between ICA angles and time to first maneuver, no significant correlation was found (Pearson’s r = −0.09, p = 0.097). For groin puncture to first maneuver, the optimal threshold was found at 89° (accuracy 0.61) and for groin puncture to final TICI was found at 92° (accuracy 0.55). Based on these findings, the best overall discriminatory power was derived for an angle of 90°, which was further used for classification.

Of the 758 ICAs assessed, 357 ICAs were classified manually as having an acute inner angle of ≤ 90° and 335 as having an obtuse inner angle of > 90°, while 66 were not assessable, of which 32 were on the occluded side. After exclusion of the cases with tandem occlusions, 316 occluded side ICAs were tested for association with procedural times.

Testing for normal distribution was negative, therefore the Wilcoxon-Mann-Whitney test was used to explore statistical significance. The angle of the ICA ipsilateral to the occlusion was strongly associated with the overall procedure time (i.e. groin puncture to first thrombectomy maneuver) with a p = 0.001. The angle of the ICA ipsilateral to the occlusion was strongly associated with the EVT duration as well as the interval from groin puncture to first maneuver, both with a p = 0.001. Table 6 and Fig. 6 summarize the results of the ICA angle classification.

Both left and right ICAs of the 379 patients were used for the angle classification task. As 66 ICAs were not assessable, this resulted in 587 for the train set and 105 ICAs for the test set. Table 7 gives an overview of the train/test distribution.

Prediction of the 105 test ICAs by the trained networks took 18 s, 0.2 s per ICA. The evaluation against the test set of the trained CNN resulted in an AUC of 0.92 and an accuracy of 0.85. Figure 7 shows the evaluation against the test set for the ICA angle classification task. The online only supplemental Figure S1 shows the training progress of the fivefold cross-validation.

Of the test ICAs 49 were on the occluded side with 2 of them being tandem occlusions, resulting in 47 ICAs that were evaluated for procedural times. Testing for significance regarding the predicted ICA angles on the test set against EVT duration (p = 0.10) and groin puncture to first maneuver (p = 0.29) was negative although descriptive statistics revealed differences, especially for EVT duration (median 44.0 min for ICA angle > 90° and 75.5 min for ICA angle ≤ 90°, see Table 8; Fig. 8).