Evaluation of an automated intracranial aneurysm detection and rupture analysis approach using cascade detection and classification networks
Introduction
According to the epidemiological studies, the incidence of intracranial aneurysm is about 3.7–6 %, and the annual rupture rate for intracranial aneurysms is about 0.95 %, which accounts for more than 80 % of nontraumatic life-threatening subarachnoid hemorrhages (SAH) with high mortality and morbidity rates (Keedy, 2006). Early detection and management of aneurysms is critical in preventing intracranial hemorrhage (Keedy, 2006, Thompson et al., 2015). Due to the extensive use of advanced imaging techniques, intracranial aneurysms are being diagnosed mostly from computed tomography angiography (CTA), magnetic resonance angiography (MRA), and digital subtraction angiography (DSA). DSA is considered to be the gold standard for diagnosing intracranial aneurysms and has been extensively used for guiding interventional operations, but it is invasive with complicated procedures. On the other hand, although MRA has a high sensitivity for intracranial aneurysm detection, it usually requires more scanning time and has contraindications including pacemakers. Compared to DSA and MRA, CTA is convenient and widely available in clinical practice (Abe et al., 2002, Sichtermann et al., 2018).
CTA scanning is a non-invasive, convenient and reliable diagnostic method, but image reading not only is time-consuming but also requires special expertise. Growing workload of radiologists often leads to low sensitivity or high false-negative results for aneurysms. Besides, various objective factors such as sizes of aneurysms, hardware conditions, differences in image acquisition protocols, and image quality can affect the accuracy of intracranial aneurysm diagnosis (Westerlaan et al., 2011). Using computer-assisted diagnosis algorithms, many tools for vessel segmentation and aneurysm detection have been developed, and it is until recently deep learning algorithms have been reported for aneurysm detection both in CTA (Assis et al., 2021, Shi et al., 2020, Yang et al., 2021) and MRA (Joo et al., 2020, Miki et al., 2016, Terasaki et al., 2021, Ueda et al., 2019). Compared to traditional image processing, deep learning-based algorithms use a series of convolutional layers to detect or segment aneurysms by optimizing the model parameters and fitting a large number of image data and can yield better detection performance than traditional image detection algorithms.
In the literature, there are different network structures to achieve aneurysm detection tasks. In (Yang et al., 2021), a convolutional neural network (CNN) was designed based on the V-Net structure for aneurysms detection, and a region proposal network (RPN) was used to generate proposals of bounding boxes with different sizes. ResNet structures (He et al., 2016) and attention modules were used for the convolutional layers. Dense atrous convolution and residual multi-kernel pooling blocks were used between encoding and decoding stages. Results showed that the network achieved high sensitivity but rendered high false-positives (e.g., 13.6 positives per scan).
Another technique for aneurysm detection is segmentation-based. In Shi et al. (2020), an end-to-end 3D CNN was proposed for aneurysm segmentation. The network structure is based on a typical encoder-decoder V-Net architecture. The difference is that for the skip connection at the highest resolution, a residual block and a dual attention block are used to focus more on this resolution for deep convolutional processing, and the attention blocks can help learn long-range contextual information. The results showed that both patient-level and lesion-level sensitivity were improved compared to those of radiologists and expert neurosurgeons. Notice that segmentation algorithms require annotated masks of aneurysms, and to simplify annotation, each aneurysm was approximated by a sphere with associated data augmentation and sampling in Assis et al. (2021).
Similar network structures were also used for aneurysm detection/segmentation in MRA. In Ueda et al. (2019) trained the ResNet-18 (He et al., 2016) from scratch for aneurysm detection and obtained sensitivities of 90–92.5 % with 5 false-positives. Similarly, in Joo et al. (2020), Joo et al. used a 3D ResNet, the sensitivity, the positive predictive value, and the specificity at subject level were 87.1 %, 92.8 %, and 92.0 % for the internal testing dataset (147 aneurysms, 120 positive, and 50 negative subjects) and 85.7 %, 91.5 %, and 98.0 % for the external testing dataset (63 aneurysms, 56 positive, and 50 negative subjects), respectively. In Terasaki et al. (2021), a multi-dimensional CNN (MD-CNN) was applied to process a 2D (maximal intensity projection, MIP) volume patch and a 3D patch, and then they were combined after global average pooling of the feature maps for aneurysm detection using fully connected layers in the final step. Results on external testing data showed that the highest sensitivity of MD-CNN was 89.1 % with 4.2 false-positives per subject.
In summary, the popular ResNet (He et al., 2016) structure can be used as a classifier to identify whether an input image volume patch consists of an aneurysm or not. The V-Net structure has been widely used for segmentation-based aneurysm detection, wherein candidates are first segmented before making the decision, among which the RPN structures can make multiple region proposals to fit different shapes and sizes of aneurysms. Results showed that RPN performed better than original V-Net.
However, high false-positive rates still necessitate radiologists to rule out them after automated processing. Therefore, it is important to study the best network configuration to achieve low false positives while maintaining high sensitivity. To deal with these challenges, in this paper, we propose and evaluate a pipeline using cascade detection and classification models. Specifically, feature pyramid network (FPN) is first used for aneurysm detection with high sensitivity, and then a ResNet aneurysm classifier (CLS) is trained to rule out false positives. In order to fully exploit the vesselness information rather than implicitly relying on convolutions to extract such image features, we apply Hessian filters and calculate the vesselness map as an additional input channel to the network. Finally, as ruptured aneurysms are life-threatening, we studied radiomics features and compared machine learning-based and deep learning-based rupture classification models.
In experiments, we used head and neck CTA images of 1508 subjects from two medical institutions for training and testing the proposed pipeline. We compared the performance of a) FPN detector; b) FPN with classification cascade model (FPN+CLS); and c) FPN with classification using vesselness channel (FPN+VCLS). For rupture classification, radiomics features were extracted from the volume patches, followed by LASSO feature selection, and support vector machine (SVM) or linear regression (LR). A ResNet rupture classification network was also evaluated. Finally, ResNet features and radiomics features are combined to form a SVM classification for comparison with other classifiers. The results indicated that using vesselness information in the cascade detection can improve the detection performance.It is also feasible to assist rupture analysis using the detection, segmentation, and classification pipeline.
Section snippets
Methods
The CTA aneurysm analysis pipeline consists of three steps, namely detection, segmentation and rupture classification. Fig. 1 illustrates the pipeline for aneurysm analysis in this work. We mainly focus on aneurysm detection and rupture classification in 2.1 Cascade FPN aneurysm detection, 2.2 Multi-channel aneurysm classification network and study whether a cascade structure including a 3D FPN and a classification network using vesselness enhancement channel can improve the performance of
Dataset and image preprocessing
Table 1 lists the sample images used in this study. The images are from two medical institutions, and altogether 1508 images of different subjects were collected. CTA images were acquired with resolution ranging from 0.4 mm to 0.7 mm in axial plane and 0.6–1.25 mm in slice distances from various CT devices (GE, Siemens). Only images with intracranial aneurysms were selected, and they were captured either from routine examination or from clinics after rupture (CTA images captured within one day
Discussion
The objectives of this paper were to evaluate the performance of an aneurysm detection, segmentation and classification pipeline, and its feasibility for clinical use with relatively large number of data. The network model was built upon FPN, which has been proved to be successful in lesion detection. In practice, because FPN proposes multiple candidates that could be overlapping each other and also results in large false positives, we merged the output candidates and then used a cascade ResNet
Conclusion
We proposed an aneurysm detection and analysis pipeline based on a cascade model of FPN and ResNet aneurysm classification. To better learning and distinguishing vessels from aneurysms, a multi-channel input that takes the original and vesselness image patches is utilized to boost performance for the FPN-classification networks. After detection, aneurysms within the detected image patches are segmented, and their radiomics features are extracted to study the differences between ruptured and
CRediT authorship contribution statement
Ke Wu: Conceptualization, Investigation, Data curation, Writing – review & editing. Dongdong Gu: Methodology, Validation, Writing – original draft. Peihong Qi: Resources, Data curation. Xiaohuan Cao: Writing – review & editing. Dijia Wu: Writing – review & editing. Lei Chen: Supervision. Guoxiang Qu: Investigation. Jiayu Wang: Investigation. Xianpan Pan: Investigation. Xuechun Wang: Investigation, Validation. Yuntian Chen: Resources, Data curation. Lizhou Chen: Resources, Data curation. Zhong
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This paper was supported by the National Key Research and Development Program of China (No.2018YFC0116400) and the National Natural Science Foundation of China (No.81901708 to J.H.L.).
References (35)
- et al.
An artificial-intelligence lung imaging analysis system (ALIAS) for population-based nodule computing in CT scans
Comput. Med Imaging Graph
(2021) - et al.
Automated detection of intracranial aneurysms based on parent vessel 3D analysis
Med. Image Anal.
(2010) - et al.
Clinical benefits of rotational 3D angiography in endovascular treatment of ruptured cerebral aneurysm
AJNR Am. J. Neuroradiol.
(2002) - Assis Y., Liao L., Pierre F., Anxionnat R., Kerrien E., 2021. An efficient data strategy for the detection of brain...
- Bizjak, Ž., Likar, B., Pernuš, F., Špiclin, Ž., 2020. Vascular Surface Segmentation for Intracranial Aneurysm Isolation...
- Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S., 2020. End-to-end object detection with...
- et al.
Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach
Biometrics
(1988) - Dosovitskiy A., Beyer L., Kolesnikov A., Weissenborn D., Zhai X., Unterthiner T., et al., 2020. An image is worth 16×16...
- Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A., 1998. Multiscale vessel enhancement filtering....
- et al.
Computational radiomics system to decode the radiographic phenotype
Cancer Res.
(2017)
Comparison of 2D and 3D U-Nets for Organ Segmentation. Auto-Segmentation for Radiation Oncology
A system to detect cerebral aneurysms in multimodality angiographic data sets
Med. Phys.
A deep learning algorithm may automate intracranial aneurysm detection on MR angiography with high diagnostic performance
Eur. Radiol.
An overview of intracranial aneurysms
Mcgill J. Med.
Clinical evaluation of AI software for rib fracture detection and its impact on junior radiologist performance
Acta Radiol.
Cited by (0)
- 1
K. Wu and D. Gu: equal contributors.