Abstract
Stroke is the second leading cause of death globally and the first leading cause of mortality in China. 87% of strokes is the ischemic strokes which are mostly caused by the cerebrovascular blockage. The cerebrovascular blockage is resulted from an embolus or plaque rupture. A major source of emboli is carotid atherosclerosis which develops inside the carotid artery wall and will narrow the artery lumen. In addition, ruptured plaques may flow and block the downstream small vessels. To monitor the carotid atherosclerosis and evaluate the treatment effect, the sensitive biomarkers are required. In particular, the previous studies showed that vessel-wall-volume (VWV) and localized vessel-wall-thickness (VWT) measured from three dimensional ultrasound (3DUS) carotid images are sensitive to anti-atherosclerotic effects of medical/dietary treatments. VWV and VWT measurements require the lumen-intima (LIB) and media-adventitia boundaries (MAB) for the common and internal carotid arteries (CCA and ICA). However, manual segmentation is time-consuming and has high intra-individual variability. In addition, most existing segmentation techniques could only segment the CCA. An approach able to delineate the MAB and LIB of the CCA and ICA was required to accelerate VWV and VWT quantification. Compared to 3DUS, 3D black-blood magnetic resonance imaging (MRI) has higher resolution, which could provide a better characterization of vessel wall pathology. This thesis presents three methods to perform carotid segmentation on 3DUS and 3D black-blood MRI.In the second chapter, we introduce a carotid segmentation method based on 2D convolution neural network (CNN). In this study, we conduct the segmentation for CCA and ICA independently using the proposed two-channel U-Net, which was trained by a novel adaptive triple Dice loss (ADTL) function. The segmentation of the ICA is more difficult than CCA since its appearance is similar to the nearby ECA and it has a smaller diameter than CCA. Therefore, for ICA segmentation, two bounding boxes were manually identified on the closest and furthest ICA axial slices from the bifurcation. The bounding boxes for the intermediate slices were automatically generated by linear interpolation. The data augmentation is performed on the training dataset by interpolating manual segmentation along the longitudinal direction to enlarge the training set. A test-time augmentation (TTA) approach was also applied to improve the segmentation accuracy, where segmentation was performed three times independently using the same model based on the original axial images and its transversely and vertically flipped versions; the final segmentation was obtained by pixel-wise majority voting of the three outputs generated in the last step. Experiments with 224 3DUS volumes showed that our method achieved a Dice similarity coefficient (DSC) of 95.1%±4.1% and 91.6%±6.6% for the MAB and LIB, in the CCA, respectively, and 94.2%±3.3% and 89.0%±8.1% for the MAB and LIB, in the ICA, respectively. The results also indicated that TTA and ATDL could independently produce a statistically significant improvement on all boundaries delineation accuracy except the LIB in ICA.
In the third chapter, we present a carotid segmentation from 3DUS based on a two-stage segmentation framework. The framework is composed of two 3D CNNs, which are centerline extraction network (CHG-Net) and dual-stream centerline guided network (DSCG-Net). Correctly locating artery is crucial for accurate segmentation of the carotid since the carotid is hard to detect especially for ICA. This issue was tackled by takine advantage of the arterial centerline to improve the localization accuracy of the segmentation network. We developed the CHG-Net to produce a heatmap indicating high probability regions of the arterial centerline for each 3DUS volume. Then the heatmap was fused with the 3DUS image by the DSCG-Net to delineate the MAB and LIB for CCA and ICA. The DSCG-Net utilizes a channel and a spatial attention mechanism to fuse multi-scale features extracted by the dual-stream encoder, and a centerline heatmap reconstruction side branch connected to the end of the encoder to increase the segmentation ability of the network when the heatmap is not accurate enough. Experiments on 224 3DUS volumes produce a DSC of 95.8±1.9% and 92.3±5.4% for CCA MAB and LIB, respectively, and 93.2±4.4% and 89.0±10.0% for ICA MAB and LIB, respectively. In addition, our approach outperformed four state-of-the-art 3D CNN segmentation models and these models boosted by centerline heatmap integration. To avoid the mislocation of ICA, the user intervention is also allowed. The user could locate the centerpoint of the distal ICA axial slice and generate the modified centerline heatmap. This user intervention could evidently improve the ICA segmentation performance.
In the fourth chapter, we focus on a two-stage segmentation framework from 3D black-blood MRI. A novel carotid black-blood MRI approach with short scanning time, 3D Motion Sensitized Driven Equilibrium prepared Rapid Gradient Echo (3D-MERGE) was previously developed to visualize the arteries with higher longitudinal dimension. We propose a mixed dimension segmentation framework consisting of a 3D and a 2D CNNs to segment MAB and LIB for CCA, ICA and ECA from 3D black-blood MRI. Our method firstly delineates MAB coarsely using a 3D Multiscale U-Net and then conducts MAB and LIB segmentation using a ROI U-Net with the 2D MRI axial slices masked by the expanded MAB produced by the 3D Multiscale U-Net as input. Compared to the previous methods, our method does not require any user intervention such as manual ROI identification in the inference. In addition, we also developed a method to produce the additional pseudo segmentation masks to help improving the segmentation performance on sparsely annotated 3D segmentation dataset. We tested our approach on Carotid Artery Vessel Wall Segmentation Challenge dataset to evaluate the effectiveness of proposed pseudo masks generation method. Also, our approach outperformed the top one solution in the benchmark of Carotid Artery Vessel Wall Segmentation Challenge and several state-of-the-art segmentation models.
In this thesis, we presented the deep learning models for the carotid segmentation from 3D US and 3D black-blood MRI. Compared to the previous carotid segmentation approaches, our method required less user interaction and achieved comparable accuracy by involving the novel segmentation loss and centerline information. In addition, the proposed surrogate ground truth generation algorithm could produce the segmentation masks for the unsegmented slices. With the generated surrogate masks, our model could achieve high accuracy with a small number of annotated slices in each 3D MRI volume.
| Date of Award | 20 Sept 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Kwok Leung CHAN (Supervisor) & Chi Yuen Bernard CHIU (External Co-Supervisor) |