Artificial Intelligence–Based 3D Angiography for Visualization of Complex Cerebrovascular Pathologies

Patient Selection

In a retrospective analysis, 3D-DSA data sets from 28 patients (age_mean = 58.7 [SD, 12.56] years; n_female/male = 11:17) with unruptured AVMs (n_AVM = 10), unruptured CAs (n_CA = 10), and dural arteriovenous fistulas (n_dAVF = 10) were evaluated. All data sets were acquired before treatment of the underlying condition in the ICA in 17 cases (57%), the vertebral artery in 9 cases (30%), and the external carotid artery in 4 cases (13%). Informed consent was obtained from all patients enrolled.

3D Angiography

The prototypical 3D angiography (3DA) refers to a novel postprocessing approach with the aim of generating DSA-like 3D volumes with diagnostic image quality. In contrast to 3D-DSA, the AI-based 3DA requires only a single contrast-enhanced run (fill run) to differentiate vessels from adjacent structures.

From a technical perspective, the core of the 3DA technique can be seen as a 3D segmentation task and is based on a classification of contrast-enhanced vascular and nonvascular structures in fill run data by applying a deep convolutional neural network¹⁹ specifically trained for this task before our evaluation. In particular, the convolutional neural network is trained to determine a 3D binary segmentation mask on the basis of the fill run data, which can, in turn, be applied to the fill run data to generate a DSA-like 3D volume. The chosen network architecture is conceived as a fast 3D network with a small memory footprint. Thus, a feed-forward 3D convolutional neural network using rectified linear unit activations normalized through batch normalization²⁰ was selected. Due to the binary nature of the classification task (ie, distinguishing vascular and nonvascular structures), the binary cross-entropy was set as a loss function. The optimization is based on the Adam optimization approach,²¹ an extension of the classic stochastic gradient descent, using a batch size of 64. The base learning rate was set to 0.01. To incorporate not only local but also surrounding contextual information into the classification, we used dilated convolutions.

Analogous to an application in 1D scenarios such as generating audio waveforms (as proposed in van den Oord et al²²), the concept of dilated convolutions can be applied to multidimensional application scenarios²³ as well. The network features 4 convolutional layers with 8 filters each, a kernel size of 5 × 5 × 5 voxels, and a dilation rate of 2. The final layer features a kernel size of 3 × 3 × 3 voxels with 1 result filter that constitutes the classification result. The advantage of this formulation is that the proposed algorithm is fully convolutional and allows a volume of arbitrary size (eg, 512 × 512 × 512 voxels) to be used as input to the classification. Furthermore, the chosen architecture allows a reasonable trade-off, especially for larger 3D volumes between the application run time and classification quality. Key aspects of the network architecture are illustrated in Fig 1.

• Download figure
• Open in new tab
• Download powerpoint

FIG 1.

Key aspects of the chosen network architecture. In addition to the in- and output of the network, kernel sizes, dilation rates, number of filters, and activation function (ie, rectified linear unit [ReLu]) are shown. Conv3D indicates 3D convolutional neural network.

In total, 98 retrospectively selected 3D-DSA data sets (free from metal or motion artifacts) covering a variety of conditions (to prevent overfitting) represented the training data for the 3DA algorithm. The training process was based on the presentation of both extracted contrast-enhanced runs (fill runs) and complete 3D-DSA volumes (containing mask and fill runs) to the deep convolutional neural network, aiming at a suitable classification between contrast-enhanced vascular and nonvascular structures, ie, the segmentation of the vascular structures in a 3D volume. For each individual data set used for the training of the 3DA algorithm, vascular extraction (ie, the derivation of a 3D binary segmentation mask) was performed by thresholding the subtracted 3D-DSA volumes. To determine a reasonable threshold, we considered the filling of the vasculature as well as the exclusion of undesirable influences (such as inconsistencies, image artifacts, noise, and so forth). To cover a sufficient variety of potentially suitable thresholds (eg, comparable with the personal preferences of multiple different clinical experts for a specific volume), we also included multiple thresholding choices simultaneously in the training process. Finally, this form of augmentation is intended to provide a better overall performance of the 3DA algorithm.

For the classification, randomly selected patches with a size of 31 × 31 × 31 voxels were applied from the presented 3D volumes during the training. The chosen size of the patches is the minimum input size for the network and results in a single-voxel response. The motivation for this choice was to foster randomness during the training, which, in turn, can contribute to the robustness and quality of the resulting classifier. All parameters were determined using hyperparameter tuning. Note that during the application of the prototypical 3DA method, the processing was not patch-based but was performed on an entire volume data set. Subsequent multimodal testing and validation of the prototypical algorithm of this classification were performed with separate 3D-DSA data sets. No data set of the training or validation was part of our series for evaluation of 3DA in cases of AVMs, CAs, and dAVFs.

Data Acquisition and Postprocessing

3D-DSA was acquired using a biplane flat panel detector angiographic system (Artis zee biplane; Siemens). By means of standard angiographic methods, a 5F catheter was positioned in the proximal ICA, external carotid artery, or vertebral artery to obtain 3D-DSA data sets using the standard acquisition protocol (5-second 3D-DSA) as provided by the manufacturer. Here, an initial rotational scan (native mask run) is followed by a second rotational scan (contrast-enhanced fill run) of 5 seconds each. Each run yields 133 projections (rotational angle = 200°). The detector dose per projection image is selected as 0.36 µGy (70 kV, 1240 × 960 detector elements with 2 × 2 binning of pixels, projection on a 30 × 40 cm flat panel size, with increment of 1.5°/frame, a frame-rate of 30 frames/s). According to our protocol, a manual injection of contrast was initiated 1 second before the beginning of the fill run, maintained for 6 seconds, and stopped after the C-arm system had covered the complete rotation angle. The total contrast volume was 15 mL of Imeron 300 (iopamidol; Bracco).

Both mask and fill runs of the acquired 3D-DSA data sets were transferred to a dedicated workstation (syngo X Workplace; Siemens) running both commercially available software for conventional 3D-DSA postprocessing and an additional software prototype plug-in for the 3DA postprocessing. Reconstruction of the 3D-DSA volumes was performed using data derived from both runs, whereas reconstruction of 3DA volumes was accomplished using only data from the fill runs.

According to standardization, we used conventional reconstruction parameters for both 3DA and 3D-DSA (kernel type: edge-enhanced; characteristics: smooth; 512 × 512 image matrix).

Image Evaluation

The image quality of all data sets was evaluated for parameters that could compromise the diagnostic value using a 5-fold grading-scale (Table 1). The 3D-DSA and 3DA reconstructions were assessed in a consensus reading by 2 experienced neuroradiologists (6 and 10 years of clinical experience) blinded to the type of reconstruction (based on either the subtraction technique or the AI algorithm).

Assessment of 3DA and 3D-DSA Reconstructions

Vessel Geometry Index.

For all 3DA- and 3D-DSA data sets, the maximum transversal diameters of the injection vessels (in millimeters; vessel diameter [VD]1/VD2) were measured in multiplanar reconstructions. Respecting the three-dimensionality of these data sets, the ratio of VD1 and VD2 was defined as the vessel geometry index (VGI = VD1/VD2).¹⁸ Measurements of the anterior and posterior circulation were performed at the C4 segment of the ICA and the V4 segment of the vertebral artery, respectively. If branches of the external carotid artery were involved, the maximum transversal diameter of the proximal external carotid artery was measured. See Fig 2 for an illustrative measurement of VD1 and VD2 at the ICA.

• Download figure
• Open in new tab
• Download powerpoint

FIG 2.

Sample measurement of the vascular dimensions of an ICA (C4 segment) in a 3D-DSA data set. The maximum vessel diameters have been assessed for both 3D-DSA and 3DA in 2 projections each (VD1 and VD2) using multiplanar reconstructions. A, Sagittal orientation of the C4 segment. B, Coronary orientation of the C4 segment. C, Finally, both vessel diameters allow the calculation of the VGI (VGI = VD1/VD2).

Statistical Analysis

Statistical analysis was performed using commercially available software (SPSS Statistics, Version 20; IBM).

Because of their noncontinuous character, qualitative parameters (eg, location of the pathology and so forth) were analyzed using descriptive statistics.

Quantitative parameters from both groups (eg, vessel diameters, VGIs, diameter of the AVM nidus, diameter of the main dAVF feeder, and diameter of aneurysmal sac) were tested for normal distribution using the D'Agostino-Pearson test (if P > .05, normality was accepted) and were compared using the Pearson correlation coefficient (r) and a paired t test (P), respectively. Any other quantitative parameters (eg, number of AVM feeders, aneurysms on AVM feeding arteries, number of intranidal aneurysms, number of aneurysmal blebs, and so forth) were analyzed using descriptive statistics.