Development of a cerebral aneurysm segmentation method to prevent sentinel hemorrhage

The sentinel clot sign is a useful computed tomography finding for the evaluation of probable anatomic sites of hemorrhage. Sentinel leaks produce sudden focal or generalized head pain that may be severe. Sentinel headaches precede the aneurysm rupture. To be further specific, cerebral Aneurysm (CA) is an abnormal inflammation within the brain blood vessels that is commonly detected after the age around forty (Higashida 2003; Nikravanshalmani et al. 2013). The mortality rate of this disease is between 30% and 40%, whereas the morbidity rate ranges between 35% and 60% (Li and Li 2010). Computer Aided Diagnosis (CAD) algorithms have been helpful in reducing the bias, time consumption, increasing the CA detection accuracy, etc. Wong (2005). Medical image segmentation (MIS) algorithms are important in the diagnosis phase.

There have been several methods in image segmentation that include both traditional methods and machine-learning/deep-learning methods. In deep learning (DL)-based methods, the neural networks (NN) used for image segmentation, have received much attention due to their end-to-end nature and state-of-the-art performance. Deep neural networks gained more popularity after their success at ImageNet Large Scale Visual Recognition Challenge 2012 (Russakovsky et al. 2015). With the improvement in GPU architectures and the development of dedicated deep learning libraries [e.g., Tensorflow (Abadi et al. 2016) and PyTorch (Paszke et al. 2019)], neural networks have been in use for a wide range of computationally heavy computer vision tasks (e.g., object recognition and segmentation). For image segmentation, the basic schematic diagram of a convolutional neural networks (CNNs) is given in Fig. 1; it comprises a series of convolution and pooling layers to condense the input image into a dense feature representation. The feature representation is utilized to reconstruct the segmentation mask using deconvolution and upsampling layers. During NN training, preprocessed images (e.g., normalized images with contrast enhancement) are fed to the network to generate the segmentation probabilities (forward propagation). A loss function computes the discrepancy between the NN prediction and the ground truth. Finally, the weights of the network are updated using an optimizer (e.g., Adam Kingma and Ba 2014) in the back-propagation phase. Altogether, the type preprocessing, architecture, choice of the loss function and optimizer, determine the segmentation accuracy and inference time of the NNs.

Over the years, although several CNNs (Khan et al. 2021; Siddique et al. 2021; Torres-Velazquez et al. 2021) have been proposed, the neural networks have faced several challenges too. For instance, the neural networks are susceptible to variations in data like changes, in contrast, noise, and resolution (Thambawita et al. 2021; Orlando et al. 2022). In a clinical setting, these variations are expected in the data due to multiple machines with several acquisition parameters that can cause the data distribution to change. One technical limitation for training the neural networks is due to the limited quantity of clinical data, resulting in overfitting (i.e., poor generalizability). Additionally, the training procedure of the neural networks does not provide any convergence guarantees. Other technical challenges include the black-box nature of deep learning-based approaches, which downplays the reliability of the neural networks in clinically sensitive settings. Thus in this paper, we propose a traditional/conventional algorithm to segment the CA from 3DRA datasets using a multiresolution statistical approach, where the Contourlet Transform (CT), in conjunction with Hidden Markov Random Field, is used to segment 2D images in the Contourlet domain.

There is rich literature in conventional MIS that includes semi-automatic and automatic (Dakua 2013; Dakua and Sahambi 2011). Broadly, the semi-automatic approaches include threshold-based (Mitra and Chandra 2013), model-based (Dakua et al. 2019), graph-based (Dakua and Sahambi 2010, 2011) methods that need human intervention (Chen et al. 2014), which is tedious and prone to operator variability; these factors certainly affect the segmentation accuracy. Therefore, automatic approaches were introduced. In Mitra and Chandra (2013), spatial filtering and dual global thresholding are considered, where three parameters (i.e. two threshold values and filter mask size) are user-defined. Yang et al. present a dot enhancement filter that includes thresholding, and a region growing technique. However, it lacks the inconsistency of the sensitivity, where it ranges between 80% and 95% for CAs larger than 5 mm and between 71% and 91% otherwise. Bogunovic et al. (2011) present a Geodesic Active Regions (GAR) based algorithm. However, there is the possibility of merging either a vessel with a CA or two vessels together. This partly happens either because of the insufficiency of the imaging resolution and the low blood flow in the small vessels. In Hentschke et al. (2011), a blobness filter and a k-means technique are adapted. The algorithm results in a false positive rate reaching 20.8%. Jerman et al. (2015), present a blobness enhancement filter combined with Random Forest Decision (RFD) classifier and a grow-cut technique, where it is assumed that a saccular CA consists of more than 15 voxels. Thus, this assumption reduces the chances to detect small CAs. Suniaga et al. (2012), propose a fuzzy logic in a level set framework along with a SVM classifier to segment saccular CAs. In Zhang et al. (2012), a method on Conditional Random Field (CRF) and gentle Adaboost classifier is proposed that is trained on six datasets and tested on two datasets. However, machine Learning (specifically, Deep Learning (DL)) models need extensive data to capture the underlying distribution for generalization in real-world systems. Additionally, the black box nature of the DL models adds uncertainty to their predictions, thus limiting their use as a second opinion in clinical setting (Su et al. 2019). Lu et al. (2016) present a method, where multi-scale filtering is used to enhance the vessels suppressing the noise and a mixture model is built to fit the histogram curve of the preprocessed data. Finally, expectation maximization is used for parameter estimation. Kai et al. (Lawonn et al. 2019) present a graph-cut based method for aneurysm segmentation. Yu et al. (Chen et al. 2020) present a geodesic active contour (GAC) and Euler’s elastica model-based method for aneurysm segmentation, where GAC segments the giant aneurysms in high contrast regions and the elastica model estimates the missing boundaries in low contrast regions. Thus, the numerous automatic segmentation algorithms reported so far in the literature have potential limitations that warrant further research on automatic methods. However, the automatic methods are difficult to be controlled, as usually desired by the clinicians during their treatment planning to explore various options. Thus, this paper proposes a semi-automatic algorithm combining Contourlet Transform (CT) and Hidden Markov Random Field with Expectation Maximization (HMRF-EM) to segment CA regardless the CA shape or size.

The rest of this paper is organized as follows: in Sect. 3, the foundation and mathematical background for the proposed algorithm are presented. In Sect. 4, the proposed algorithm is discussed in detail. In Sect. 5, the objective and subjective evaluation of the proposed work are presented along with the dataset description and the environmental setup for the implementation. Section 6 includes discussion and some future work, whereas the Sect. 7 concludes the paper.

The proposed CA segmentation algorithm is illustrated in Fig. 4.

The algorithm starts by feeding the input 2D scans after enhancement. The selection of the Region of Interest (ROI) from the entire cerebral vasculature is done manually. Subsequently, the below steps are performed on each 2D image.

During the first step, CT is applied to extract features from the input image by decomposing it into six pyramidal levels and different directions for each level, where the number of directional decomposition at each pyramidal level (from coarse to fine) are: \(2^{2}\), \(2^{2}\), \(4^{2}\), \(4^{2}\), \(8^{2}\), and \(8^{2}\) (Hiremath et al. 2010; Po and Do 2006). As discussed earlier, CT consists of two main filters, LP and DFB. A ladder filter, known as PKVA filter (Phoong et al. 1995), is selected for the first filter. The PKVA filter is effective to localize edge direction as it reduces the inter-direction mutual information (Po and Do 2006). As for the second filter, 9-7 bi-orthogonal Daubechies filter (Cohen and Daubechies 1993) is selected; this filter significantly reduces the inter-scale, inter-location, and inter-direction mutual information of the Contourlet (Po and Do 2006). After the decomposition by CT, the lowpass subband image, \(a_{j}[n]\), is used to perform the rest of the steps.

To apply the second step (HMRF-EM) of the segmentation algorithm, two prior steps need to be performed. First, an initial segmentation is generated using the k-means method, which is known with its under-estimation to complement the HMRF-EM framework (AlZubi et al. 2011). In addition, a Canny edge filter (Gonzalez and Wood 2002) is applied to highlight the image edges. After that, the HMRF-EM algorithm iterates between the E-Step and M-Step to enhance the initial segmentation constrained by the Canny edge detector.

The HMRF-EM method starts after getting the initial segmentation, \(\hat{x}^{(0)}\), and initial parameters, \(\theta ^{(0)}\), obtained by the k-means clustering, the constrained image, \(ce_{j}[n]\), obtained by the Canny edge operator, and the lowpass subband image, \(a_{j}[n]\), obtained by the Contourlet decomposition. During this step, the algorithm iterates between the steps E and M to refine the initial segmented image, constrained by the Canny segmented image, resulting in the final segmented 2D image by minimizing the posterior energy function as explained in Sect. 3.3.

As the last step, Inverse Contourlet Transform (ICT) is applied to reconstruct the image. Here, the lowpass subband image, \(a_{j}[n]\), the coarsest Contourlet coefficients, is replaced by the final segmented image, \(\hat{x}^{(MAP_{\_}itr)}\). The ICT is achieved using the same filters as in the decomposition stage, where the 9-7 and PKVA filters are used for the LP and DFB, respectively.

After completing these two phases, the reconstruction of all the segmented 2D images is performed to get the final segmented 3D volume of the ROI. The pseudocode for the overall proposed CA segmentation algorithm is presented in Algorithm 3.

Both the qualitative and quantitative results are promising when tested on these 3DRA datasets. In the quantitative evaluation, the average values of accuracy, DSC, FPR, FNR, specificity, and sensitivity, are \(99.72\%\), \(93.52\%\), \(0.07\%\), \(5.23\%\), \(94.77\%\), and \(99.96\%\), respectively; an average of 4.14 over 5 is obtained in the qualitative evaluation.

It may be observed in Tables 2 and 3 that the last dataset has the worst results as compared to the remaining datasets in both the quantitative and qualitative evaluation. This may be due to the fact that the provided ground truth data does not involve the complete brain vessels tree and only a delineated ROI is provided, where some surrounding vessels are excluded. The same may be observed from Fig. 12.

The computational time to segment a CA volume is considerably fast. Table 4 reports the running time of the proposed segmentation algorithm for the ROI in seconds (s) and the whole volume of a subject in minutes (min). Even though the results are acceptable, the computational time can further be reduced by using Field-Programmable Gate Array (FPGAs) or Graphics Processing Unit (GPUs).

We have also compared the performance of the proposed method with some similar methods that have been published recently. The results are provided in Table 5. The results show that the performance of the proposed method is better than the others.

Additionally, we have compared (in Table 6) the proposed method with some popular DL-based methods, including Voxel-Morph (VM) (Balakrishnan et al. 2019), LT-Net (Wang et al. 2020), and Symmetric Normalization (SyN) (Avants et al. 2008). We have selected these methods because they have been regularly preferred by the research fraternity for comparison purpose. Among these, Voxel-Morph is probably the most famous method in recent years. The results show that the proposed method fairly performs as compared to the DL-based methods although the margin is not very significant. Furthermore, we have also compared the computational complexity involved in Table 7.

Sentinel headache is often assumed to portend an increased risk of delayed cerebral ischemia and aneurysm rebleeding. High rates of morbidity and mortality are associated with Sub-Arachnoid Hemorrhage (SAH) that is caused by a ruptured CA. Therefore, it is quite imperative to detect and diagnose CAs at an early stage. In this paper, we have presented a semi-automatic CA segmentation method using Contourlet transform and the hidden Markov random field model. Prior to use the method, we have denoised the input data to improve the contrast of the input images and make them suitable for segmentation. We have obtained promising results when tested on 3DRA datasets. Since, it is usually difficult to obtain publicly available original and corresponding ground truth data, we had rely on HMC data. The preparation of ground truth data needs a huge effort and thus, we have used only a small number of datasets. In future, we would like to increase the number of datasets and make the segmentation algorithm further robust. Furthermore, since this segmentation algorithm could be used to improve the cerebral aneurysm treatment planning, we also plan to take this into clinical routine after thorough testing and validation.