Deep learning-based recognition and segmentation of intracranial aneurysms under small sample size

Abstract

The manual identification and segmentation of intracranial aneurysms (IAs) involved in the 3D reconstruction procedure are labor-intensive and prone to human errors. To meet the demands for routine clinical management and large cohort studies of IAs, fast and accurate patient-specific IA reconstruction becomes a research Frontier. In this study, a deep-learning-based framework for IA identification and segmentation was developed, and the impacts of image pre-processing and convolutional neural network (CNN) architectures on the framework's performance were investigated. Three-dimensional (3D) segmentation-dedicated architectures, including 3D UNet, VNet, and 3D Res-UNet were evaluated. The dataset used in this study included 101 sets of anonymized cranial computed tomography angiography (CTA) images with 140 IA cases. After the labeling and image pre-processing, a training set and test set containing 112 and 28 IA lesions were used to train and evaluate the convolutional neural network mentioned above. The performances of three convolutional neural networks were compared in terms of training performance, segmentation performance, and segmentation efficiency using multiple quantitative metrics. All the convolutional neural networks showed a non-zero voxel-wise recall (V-Recall) at the case level. Among them, 3D UNet exhibited a better overall segmentation performance under the relatively small sample size. The automatic segmentation results based on 3D UNet reached an average V-Recall of 0.797 ± 0.140 (3.5% and 17.3% higher than that of VNet and 3D Res-UNet), as well as an average dice similarity coefficient (DSC) of 0.818 ± 0.100, which was 4.1%, and 11.7% higher than VNet and 3D Res-UNet. Moreover, the average Hausdorff distance (HD) of the 3D UNet was 3.323 ± 3.212 voxels, which was 8.3% and 17.3% lower than that of VNet and 3D Res-UNet. The three-dimensional deviation analysis results also showed that the segmentations of 3D UNet had the smallest deviation with a max distance of +1.4760/-2.3854 mm, an average distance of 0.3480 mm, a standard deviation (STD) of 0.5978 mm, a root mean square (RMS) of 0.7269 mm. In addition, the average segmentation time (AST) of the 3D UNet was 0.053s, equal to that of 3D Res-UNet and 8.62% shorter than VNet. The results from this study suggested that the proposed deep learning framework integrated with 3D UNet can provide fast and accurate IA identification and segmentation.

Keywords: SAH; automatic segmentation; convolutional neural network; deep learning; intracranial aneurysm.

FIGURE 1

Composition of IAs of different locations and sizes (A) Aneurysm location distribution (B) Aneurysm size distribution.

FIGURE 2

Data pre-processing process.

FIGURE 3

Convolutional neural network structures of (A) 3D UNet (Cicek et al., 2016) (B) VNet (Milletari et al., 2016), and (C) 3D Res-UNet (Kerfoot et al., 2019).

FIGURE 4

Changes of dice loss value and DSC coefficient in the training process of three models. The changes of the Dice loss on the training set of (A) 3D UNet (C) VNet, and (E) 3D Res-UNet. The changes of the DSC on the test set of (B) 3D UNet (D) VNet, and (F) 3D Res-UNet. The red and blue curves in the figure represent the training process of images with derivation and without derivation, respectively.

FIGURE 5

Changes of dice loss value and dice coefficient in the training of three models (A) Changes of dice loss on the training set (B) Changes of DSC on the test set.

FIGURE 6

Segmentation results of three models on the test set (A) V-Recall (B) DSC (C) HD.

FIGURE 7

Visualization of segmentation results of three models. The ground-truth annotations are shown in red, and the automatic segmentations of 3D UNet, VNet, and 3D Res-UNet are shown in yellow, blue, and green, respectively.

FIGURE 8

Statistical results of DSC values of segmentation results on the test set. The proportion of segmentation results within different DSC values based on (A) 3D UNet (B) VNet, and (C) 3D Res-UNet. Group A, Group B, and Group C represented samples of DSC between 0.4 and 0.6, 0.6–0.8, and 0.8–1.0, respectively.

FIGURE 9

Comparison of IA segmentation results under different sizes (A) V-Recall (B) DSC (C) HD.

FIGURE 10

3D deviation analysis results of two IA reconstructed models based on CNN segmentation. The deviation is represented in the reference geometry. The first, second, and third columns represent 3D deviation results for 3D UNet, VNet, and 3D Res-UNet on two IA reconstructed models (upper and lower rows), respectively. The red areas show an overestimation of the reference model and the blue areas indicate an underestimation. Red: +6.000 mm deviation. Green: 0.000 mm deviation. Blue: −6.000 mm deviation.