Automated Segmentation of Intracranial Thrombus on NCCT and CTA in Patients with Acute Ischemic Stroke Using a Coarse-to-Fine Deep Learning Model

Study Participants

Patients were retrospectively selected from the Safety and Efficacy of Nerinetide in Subjects Undergoing Endovascular Thrombectomy for Stroke (ESCAPE-NA1) randomized controlled trial (ClinicalTrials.gov: NCT02930018).¹³ Study approval was obtained from the ethics board at each site and the responsible regulatory authorities. Signed informed consent was obtained from the patients or their legally authorized representatives. Inclusion criteria for the main study were the following: 1) 18 years of age or older with LVO (intracranial ICA, MCA M1, or functional M1 [proximal occlusion of all M2 branches]), 2) NIHSS score of ≥5, 3) time from last seen well <12 hours, 4) pial collateral filling of ≥50% of the ischemic MCA territory, and 5) ASPECTS ≥ 5. For this study, we included only patients with available thin-section (≤2.5 mm) baseline NCCT and CTA images (single-phase or 1 phase of multiphase CTA). We excluded patients whose imaging showed the following: 1) irremediable coregistration errors (n = 13), 2) severe motion artifacts (n = 16), or 3) beam-hardening artifacts (n = 11). A total of 499 patients were included, of whom 329 were randomly selected for the training (n = 263) and internal validation (n = 66); the remaining 170 patients (independent of the derivation cohort) were used to internally test the derived model (Fig 1).

• Download figure
• Open in new tab
• Download powerpoint

FIG 1.

Flowchart of patient inclusion.

An external validation set was also used to test the generalizability of the derived DL model. This data set comprised 83 randomly chosen patients with AIS with anterior circulation occlusions from the Precise and Rapid Assessment of Collaterals Using Multi-Phase CTA in the Triage of Patients with Acute Ischemic Stroke for IV or IA Therapy (PRoVe-IT) study.^14,15 Of the 83 patients, 34 had LVO, 28 had medium-vessel occlusion (MeVO) (M2/M3/M4 segments of the MCA or A2/A3/A4 segments of the anterior cerebral artery), and 21 had no identifiable intracranial occlusion.

Reference Standard: Manual Segmentation of Thrombi

The thin-section NCCT images were automatically coregistered onto the CTA images (using the second phase if multiphase CTA was available) using rigid-body transformation (the SimpleITK packages in Python; https://pypi.org/project/SimpleITK/),¹⁶ followed by skull-stripping.¹⁷ An expert neuroradiologist visually inspected the registration results and performed manual corrections by using the 3D Slicer software (Version 4.1, https://www.slicer.org/) when the coregistration was suboptimal.¹⁸ Four trained readers (3 neuroradiologists with >5 years’ experience in vascular imaging and 1 vascular neurologist with >10 years of stroke imaging experience) manually segmented intracranial thrombi section-by-section on NCCT images while referring to the coregistered CTA images using ITK-SNAP (http://www.itksnap.org/).¹⁹ Four readers were each responsible for one-fourth of the entire data set, and they were blinded to clinical information and follow-up imaging.

All 4 readers segmented a subset of 10 patients randomly selected from the internal test set. An expectation-maximization algorithm for simultaneous truth and performance level estimation (STAPLE) algorithm was performed to generate a computational “golden reference standard” based on the 4 experts’ segmentations, which was used to assess the variability of manual segmentations.²⁰

Deep Learning Model

A 2-stage coarse-to-fine thrombus segmentation neural network was proposed on the basis of the U-net architecture.²¹ The proposed deep learning (DL) model used a multiscale training strategy with a deep-supervision mechanism. In particular, a channel and spatial attention block was designed to make the model focus more on the salient areas on images at different scales and obtain more conducive features.²² The spatial attention module generates a spatial attention map using the interspatial relationship of features. Different from channel attention, spatial attention focuses on where an informative part is, which is complementary to channel attention. To compute the spatial attention, we apply average-pooling and max-pooling operations along the channel axis first and concatenate them to generate an efficient feature descriptor.²² On the concatenated feature descriptor, we apply a convolution layer to generate a spatial attention map, which encodes where to emphasize or suppress. In this article, the patch size for spatial transform was voxels. To make full use of multiscale context information, we used a scale-aware pyramid fusion module, in which 3 parallel dilated convolutions with different dilation rates were used to capture information at different scales.²³

The detailed architecture of the proposed model is shown in Fig 2, and the details around the model architecture and hyperparameter optimization are summarized as follows: The 3D kernel size for convolutions was 3 × 3 × 3. Feature numbers at each layer were 32, 64, 128, 256, and 320 (limited to 320 to ensure sufficient context aggregation). Batch normalization was used in the proposed network, which took a step toward reducing the internal covariate shift and, in doing so, dramatically accelerated the training of deep neural nets. The batch size of the networks was 2, to enable large patch sizes, and the leaky Rectified Linear Unit nonlinearity was implemented. Stochastic gradient descent with Nesterov momentum (µ = 0.99) was used as the optimizer. The initial learning rate, dampening, batch size, and decay weight were 0.01, 0, 2, and , respectively. The learning rate was decayed throughout the training following , where is 0.01 and is the maximum of the epoch.

• Download figure
• Open in new tab
• Download powerpoint

FIG 2.

The proposed coarse-to-fine DL architecture. A 2-stage coarse-to-fine thrombus segmentation neural network used a multiscale training strategy with a deep-supervision mechanism, which consisted of the channel and spatial attention blocks and the scale-aware pyramid fusion module.

Specifically, NCCT images after skull-stripping were downsampled from the original spacing of ( voxels) to ( voxels) and passed into the coarse model to obtain maximal contextual information and localize thrombi candidates at the first coarse stage. The output of the prediction maps at the coarse stage was upsampled to the original resolution and fed into the fine stage together with the original images to obtain more detailed segmentation. In particular, the fused features of each layer at the coarse stage were upsampled and concatenated to the features of the corresponding layer at the fine stage to use both global contextual information and local detail information obtained at coarse and fine stages.

Additionally, 5-fold cross-validation on the training data set was performed to select the optimal model at the coarse and fine stages. Spatial augmentations (rotation, scaling, low-resolution simulation, and so forth) were applied in 3D to increase the diversity of the training data. The combined Dice and cross-entropy were used as the loss function.²⁴ At each level of deep supervision, the ground truth segmentation mask was correspondingly downsampled for loss computation on the basis of the size of the feature maps. The training objective was the sum of the losses ( L) at all resolutions, . Hereby, the weights ( W) were halved with each decrease in resolution, resulting in , , and so forth, and were normalized to sum 1.

Statistical Methods

Patient demographics, including clinical and imaging variables, were compared across the derivation and internal test sets using the χ² and Student t tests as appropriate.

The proposed segmentation method was quantitatively evaluated using the spatial overlap metric of the Dice coefficient (DC) and 2 boundary distance error metrics: average symmetric surface distance (ASSD) and 95th percentile of the Hausdorff Distance (HD95),²⁵ compared with the reference standard of manual (expert) segmentation. The DC is a spatial overlap index ranging from 0 to 1, where 1 indicates a perfect overlap between the reference standard and the predicted segmentation and vice versa. The ASSD and HD95 represent the average and largest HD95 errors between the 2 surfaces derived from the segmented and reference objects in 3D space. For ASSD and HD95, 0 mm indicates perfect segmentation. The correlations regarding thrombus length and volume between the DL-derived model and manual measurements were analyzed using the Pearson correlation and Spearman correlation as appropriate to the data distribution.²⁶ The correlation was considered excellent if ≥0.70 and good if between 0.5 and 0.7.²⁶ Absolute and relative errors of thrombus volume and length were also calculated. These metrics were all calculated in 3D space at a patient level and applied onto both the internal data set (ESCAPE-NA1) and the external test data set (PRoVe-IT) for internal and external validation, respectively. On the basis of the segmentation results, the model predictions can be also used to distinguish the patients with LVO versus without LVO after localizing the segmented thrombus and thresholding thrombus volume. We further assessed the specificity and sensitivity of the DL model in classifying intracranial thrombi (LVO versus non-LVO and occlusion versus nonocclusion) in the external data set.

All statistical analyses were performed using the SciPy package (https://scipy.org/). All P values were 2-sided, and statistical significance was defined as P < .05.