Evaluation of an automated intracranial aneurysm detection and rupture analysis approach using cascade detection and classification networks

https://doi.org/10.1016/j.compmedimag.2022.102126Get rights and content

Highlights

  • Proposed an aneurysm detection and analysis pipeline based on a cascade model of FPN and ResNet classification with CTA images.

  • Taking vesselness features as additional channels to improve the performance of FPN-classification networks

  • Deep learning-based and machine learning-based methods are used to study the differences between ruptured and un-ruptured aneurysms.

  • Results showed satisfactory performance in aneurysm detection and ruptured and un-ruptured aneurysm classification.

Abstract

Intracranial aneurysm is commonly found in human brains especially for the elderly, and its rupture accounts for a high rate of subarachnoid hemorrhages. However, it is time-consuming and requires special expertise to pinpoint small aneurysms from computed tomography angiography (CTA) images. Deep learning-based detection has helped improve much efficiency but false-positives still render difficulty to be ruled out. To study the feasibility of deep learning algorithms for aneurysm analysis in clinical applications, this paper proposes a pipeline for aneurysm detection, segmentation, and rupture classification and validates its performance using CTA images of 1508 subjects. A cascade aneurysm detection model is employed by first using a fine-tuned feature pyramid network (FPN) for candidate detection and then applying a dual-channel ResNet aneurysm classifier to further reduce false positives. Detected aneurysms are then segmented by applying a traditional 3D V-Net to their image patches. Radiomics features of aneurysms are extracted after detection and segmentation. The machine-learning-based and deep learning-based rupture classification can be used to distinguish ruptured and un-ruptured ones. Experimental results show that the dual-channel ResNet aneurysm classifier utilizing image and vesselness information helps boost sensitivity of detection compared to single image channel input. Overall, the proposed pipeline can achieve a sensitivity of 90 % for 1 false positive per image, and 95 % for 2 false positives per image. For rupture classification the area under curve (AUC) of 0.906 can be achieved for the testing dataset. The results suggest feasibility of the pipeline for potential clinical use to assist radiologists in aneurysm detection and classification of ruptured and un-ruptured aneurysms.

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)

  • L. Chen et al.

    An artificial-intelligence lung imaging analysis system (ALIAS) for population-based nodule computing in CT scans

    Comput. Med Imaging Graph

    (2021)
  • A. Lauric et al.

    Automated detection of intracranial aneurysms based on parent vessel 3D analysis

    Med. Image Anal.

    (2010)
  • T. Abe 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...
  • E.R. DeLong 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....
  • J.J.M. van Griethuysen et al.

    Computational radiomics system to decode the radiographic phenotype

    Cancer Res.

    (2017)
  • D. Gu et al.

    Comparison of 2D and 3D U-Nets for Organ Segmentation. Auto-Segmentation for Radiation Oncology

    (2021)
  • He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep Residual Learning for Image Recognition. IEEE Conference on Computer...
  • C.M. Hentschke et al.

    A system to detect cerebral aneurysms in multimodality angiographic data sets

    Med. Phys.

    (2014)
  • B. Joo et al.

    A deep learning algorithm may automate intracranial aneurysm detection on MR angiography with high diagnostic performance

    Eur. Radiol.

    (2020)
  • A. Keedy

    An overview of intracranial aneurysms

    Mcgill J. Med.

    (2006)
  • Lin, T.-Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S., 2017. Feature pyramid networks for object...
  • X. Liu et al.

    Clinical evaluation of AI software for rib fracture detection and its impact on junior radiologist performance

    Acta Radiol.

    (2021)
  • Cited by (0)

    1

    K. Wu and D. Gu: equal contributors.

    View full text