Development of a diagnostic model for acute ischemic stroke early identification based on SE-ResNeXt

doi:10.21203/rs.3.rs-4869692/v1

Download PDF

Article

Development of a diagnostic model for acute ischemic stroke early identification based on SE-ResNeXt

https://doi.org/10.21203/rs.3.rs-4869692/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

Acute ischemic stroke (AIS) is a disease with high incidence rate and mortality. At present, the accuracy of AIS recognition based on Non contrast computed tomography (NCCT) images is not sufficient to meet clinical needs. We hope to develop and validate an AIS recognition model that can achieve timely and accurate recognition.

Methods

We retrospectively collected NCCT images of 287 patients from the Second Affiliated Hospital of Zhejiang University School of Medicine, and randomly divided them into a training set n = 230 and a testing set n = 57 according to a ratio of 8:2. We developed a deep learning AIS recognition model based on 3D SE-ResNeXt. The classification performance of the model was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, and F1 score.

Results

The AUC, accuracy, sensitivity, specificity, and F1 of the model on the training set were 0.96, 0.94, 0.91, 0.94, and 0.92, respectively. The AUC, accuracy, sensitivity, specificity, and F1 on the test set were 0.90, 0.88, 0.82, 0.86, and 0.84, respectively. Compared with other deep learning models, the model used in this article has the best performance.

Conclusion

These results indicate that the proposed method can achieve early identification of acute ischemic stroke on NCCT images, which has high clinical significance.

Health sciences/Medical research/Experimental models of disease

Health sciences/Diseases/Neurological disorders/Hypoxic ischaemic encephalopathy

Acute ischemic stroke (AIS)

Deep Learning

Non contrast computed tomography (NCCT)

Early Identification

SE-ResNeXt

Acute ischemic stroke (AIS) is the most common type of stroke, accounting for approximately 70% of strokes in my country [1, 2]. Imaging examinations play an important role in promptly and accurately determining the location of AIS, assessing the degree of brain injury, and determining the vascular lesions causing ischemic injury, and can provide an important basis for formulating diagnosis and treatment plans [3]. Among imaging examination methods, CT scan has become the preferred method in most medical centers due to its wide popularity, fast scanning speed, and easy detection of intracranial hemorrhage [4]. A series of research reports at home and abroad have also confirmed that Non-contrast computed tomography (NCCT) images can provide effective information for the early identification of AIS [5–6].

Non-contrast computed tomography also has the following specific problems in identifying early cerebral infarction: (1) Although some signs in Non-contrast computed tomography images, such as high blood vessel density sign, blurred lentiform nucleus, etc., are helpful in identifying early ischemic signs, its early infarction detection rate is not high. There is a systematic review in the literature, and the results show that the detection rate of early infarction signs is only 61% when CT scan is performed within 6 hours after the onset of stroke [7]. (2) Because early infarction signs on plain CT scan images may be very subtle and easily confused with normal physiological changes or senile changes, early infarction signs are often underestimated or overestimated in controlled environments [8]. (3) Signs of early cerebral infarction have low consistency among different evaluators [9].

The booming development of artificial intelligence and deep learning technology provides a new way for data analysis in the medical field [10–16]. In the study of brain diseases, deep learning algorithm models based on CT images play an important role in the diagnosis and treatment evaluation of brain diseases [17–20]. Zhang [21] et al. compared the value of MRI and CT imaging in the diagnosis of middle-aged patients with multiple cerebral infarction. On a data set of 218 cases, the total detection rate of CT examination was 56.88%. Researchers such as Phaphuangwittayakul [22] proposed an optimized deep learning framework to assist in the quantitative assessment of intracranial hemorrhage diagnosis and accurately detect different subtypes of intracranial hemorrhage through head CT scans. Experimental results show that the model is effective for three types of hemorrhage. The average accuracy reaches 96.21%. Öman O[23] et al. collected 60 cases to train a 3D convolutional neural network model, and applied the model to the study of detecting AIS from CT angiography images. The final sensitivity of diagnosis through the model was 0.93. Specificity was 0.82. van Ommen F[24] et al used dual-energy CT (DECT) to conduct statistics on 125 cases of early cerebral infarction. The results showed that non-contrast CT can distinguish normal brain parenchyma from infarcted brain parenchyma. Current research on artificial intelligence models for AIS includes relatively small amounts of data, which limits the accuracy and generalization of artificial intelligence diagnostic models.

In response to the current problems faced by AIS diagnosis, this paper proposes a diagnostic model based on SE-ResNeXt to achieve rapid and accurate diagnosis of AIS from NCCT images, which has important clinical value.

Patients

This study was approved by the Human Research Ethics Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine, and followed the guidelines of the Declaration of Helsinki. The requirement for patient informed consent was waived. Three hundred and twelve patients who received CT examinations from May 2021 to October 2023 because of suspicion of acute stroke were eligible, irrespective of symptom severity or ultimate clinical diagnosis. The exclusion criteria were as follows: (a) more than 24 hours after onset of stroke symptoms; (b) history of head injury or brain tumors; and (c) unsatisfactory image quality. The final clinical diagnosis was that documented in the patient’s hospital record during the admission by the responsible stroke-team neurologist, based on all available clinical information, including acute and follow-up brain imaging and ancillary testing. Patients with imaging evidence of cerebral infarction were given a final diagnosis of AIS. Finally, 287 patients were included in this study, and all data were randomly divided into a training set (n = 230) and a test set (n = 57) in a ratio of 8:2.

CT examinations in this study were performed using Aquilion One 320 (Canon Medical Systems), Brilliance 64 (Philips Healthcare), Optima CT540 (GE MEDICAL SYSTEMS), Sensation 16 (Siemens Healthineers), SOMATOM Definition AS (Siemens Healthineers), SOMATOM Definition Flash (Siemens Healthineers), and SOMATOM Perspective (Siemens Healthineers). The scanning parameters were set as follows: a tube voltage ranging from 100 to 130 kV, a tube current between 10 and 644 mA, and a slice thickness of 5 mm.

Image enhancement and image preprocessing

Before training the model, we preprocessed the data. As shown in Fig. 1, we first segmented the brain region using the HD-BET network[21], removing parts such as the skull and eyes. Then, we corrected the bias field of the MRI images using the N4 algorithm[22], which was implemented based on the SimpleITK's N4BiasFieldCorrection package. Finally, we resized the images to a uniform size of 512*512*18 and normalized the gray values to the range of [-0.5,0.5]. Preprocessing the images can reduce the impact of outlier cases with significant variations and enable the model to learn common features from images across different age groups.

Deep learning model development

We used the 3D SE-ResNeXt network to construct the classification model, which extends the 2D SE-ResNeXt network [25] into a three-dimensional network. The 2D and 3D versions of this series of networks are widely used in medical image analysis, including image segmentation [26], image classification [27], and lesion detection [28], etc., and have achieved good performance. The basis of the model is ResNet[29] and SENet[30]. As shown in Fig. 2, the 3D ResNeXt network draws on the split-convert-merge strategy in InceptionNet [31] to combine the residual difference module is improved to a grouped convolution extended residual block. Grouped convolution not only improves the expressive ability of the network, but also reduces the computational complexity and the number of parameters, making the network easier to train and deploy.

Model evaluation

Accuracy, sensitivity, specificity, and F1 score are used to evaluate the performance of the model on the training set and test set. Among them, sensitivity and specificity indicate the AIS category. At the same time, we drew a confusion matrix to show the model's classification performance for each category. The calculation formula of each indicator is as follows:

Where TP, FN, TN, and FP represent true positive, false negative, true negative, and false positive, respectively.

Statistical Analysis

Data processing and model construction were completed on the server in a Linux environment (Ubuntu 7.5.0). The server CPU is Intel 4215FR 3.20 GHz, with 64GB DDR4 memory. All programming uses Python 3 version 3.6.13 (Python Software Foundation). Model construction is based on the PyTorch (https://pytorch.org/) deep learning framework. The main packages used are scikit-learn module version 0.20.4, numpy module version 1.19.5, torchvision module version 0.11.2, and torch module version 1.10.1. Eight GPUs (NVIDIA RTX 3060) were used during model construction.

Distribution of data

A total of 287 cases participated in the study. There were 230 participants in the training set (mean [SD] age, 64.05 [10.67] years) and 57 participants in the validation set (mean [SD] age, 65.07 [9.92] years). According to the clinical diagnosis results, in the training set, there were 84 cases diagnosed with AIS, accounting for 36.52%. In the validation set, there were 21 cases diagnosed with AIS, accounting for 36.84%. The detailed distribution of cases in the training set and validation set is shown in Table 1.

Table 1

The distribution of cases on the training and validation sets.
	Train set (n = 230)	Test set (n = 57)	P-value
Age(years), (mean + SD)	64.05 ± 10.67	65.07 ± 9.92	0.237
Sex, n(%)			0.721
male	134(41.74)	35(38.60)
femele	96(58.26)	22(61.40)
Clinical diagnosis, n(%)			0.473
AIS	84(36.52)	21(36.84)
Non-AIS	146(63.48)	36(63.16)

Model building

After many experiments, the optimal parameters of the model were determined. SE-ResNeXt parameters are set as follows: dropout rate is 0.5, growth rate is 4, training epochs are 1000, batch size is 20, learning rate is 0.1, and the learning rate is updated every 500 trainings. In order to reduce the impact of sample imbalance on the model, during training, Viral encephalitis, Supportive encephalitis, and Non-encephalitis are assigned weights of 0.2, 0.4, and 0.4 respectively.

Model performance evaluation

The AUC of the final model on the training set and validation set was 0.96 and 0.90 respectively. As shown in Table 2, on the training set, the accuracy, sensitivity, specificity, and F1 score of the model were 0.94, 0.91, 0.94, and 0.92 respectively. On the validation set, the accuracy, sensitivity, specificity, and F1 score of the model were 0.88, 0.82, 0.86, and 0.84 respectively. Figure 3 shows the confusion matrix of the model on the training set and test set. It can be seen from the figure that most cases are divided into their corresponding groups. The above experimental results show that the model proposed in this article has high AIS identification ability.

Table 2

The performance of our model on the train set and validation set.
	Train Set (n = 230)					Validation set (n = 57)
	AUC	ACC	SEN	SPE	F1	AUC	ACC	SEN	SPE	F1
Our mode	0.96	0.94	0.91	0.94	0.92	0.90	0.88	0.82	0.86	0.84
ACC:accuracy; SEN:sensitivity; SPE:specificity.

On the dataset of this study, we trained common medical image classification models and compared the performance of each model with our method on the test set. These models include ResNet[29], DenseNet[31], ResNeXt[25], CBAM-ResNet[32], ShuffleNet[33], and Swin-transformer[34]. As shown in Table 3 and Fig. 4, our model is better than the other models in every indicator and has the highest performance.

Table 3

Performance of different models on the validation set.
Models	AUC	ACC	SEN	SPE	F1	TP	FN	FP
ResNet	0.76	0.79	0.70	0.76	0.73	16	5	7
DenseNet	0.83	0.81	0.73	0.76	0.74	16	5	6
ResNeXt	0.85	0.84	0.77	0.81	0.79	17	4	5
CBAM-ResNet	0.82	0.82	0.74	0.81	0.77	17	4	6
ShuffleNet	0.80	0.79	0.70	0.76	0.73	16	5	7
Swin-transformer	0.81	0.84	0.75	0.86	0.80	18	3	6
Our mode	0.90	0.88	0.82	0.86	0.84	18	3	4

Early identification of stroke is critical for the treatment and subsequent recovery. NCCT is a routinely employed imaging modality for stroke evaluation. However, NCCT imaging hardly shows significant changes until 12–24 hours after the onset of stroke. Our model achieved satisfactory performance in diagnosing early AIS with an AUC of 0.90, a SEN of 0.82, and a SPE of 0.86. DWI, the commonly used MR sequence for diagnosing cerebral infarction, demonstrated a sensitivity of 90% and specificity of 97% for diagnosing early AIS[35]. Compared with DWI, our model showed equivalent or possibly better sensitivity, which means that fewer patients may be missed by our model.

With the development of deep learning technology, more and more researchers are trying to use deep learning to solve challenging clinical problems. Peiyi Gao et al. [36] constructed a two-stage deep learning framework and implemented a deep learning system for segmenting and classifying 18 types of intracranial tumors from MRI data of 37,871 patients. The model also helped neuroradiologists improve the average accuracy of 1,166 patients by 12.0 percentage points. Abbasi H et al. [37] summarized the deep learning technology for automatic segmentation of cerebral ischemic stroke. Combining various studies, it can be seen that various deep learning models can effectively handle the complexity and variability of stroke lesions by learning from large-scale data sets, capturing complex patterns, and extracting meaningful features. It can effectively and accurately automatically segment ischemic stroke lesion areas from medical imaging data.

At present, there are many studies on acute ischemic stroke deep learning models combined with MRI images [38–42]. Dai et al. [43] developed a DL model based on CT angiography, which can automatically detect abnormal blood vessels in less than 2 minutes without any prior restrictions. Et al. [44] used a localization model to help locate early signs of infarction on NCCT, and were able to detect early infarction in the supratentorial and infratentorial regions, ultimately achieving an AUC of 0.7, and detecting early infarction in the supratentorial and infratentorial regions with an average area under the curve (AUC) of 0.73. The method proposed in this paper uses the SE mechanism to improve the AUC of the AIS early screening model on NCCT to 0.90 by learning 3D features, achieving a significant improvement in performance.

This study has several limitations: (1) This study is a retrospective study, so it may cause statistical bias. Prospective data need to be included in future studies. (2) This study is based on single-center data, so multi-center data also need to be included to verify the generalization ability of the model. (3) The distribution of data samples in this study is uneven, which may cause the model to focus on categories with a large number of samples. In the future, it will be necessary to collect data from more sample categories to balance the performance of the model.

We have built a deep learning model for early identification of AIS from NCCT images, which is completed using SE-ResNeXt and has high recognition accuracy. In the future, we will apply it in emergency departments to explore its practical clinical application value.

Author contributions statements

Jie Xu: Conceptualization, Formal Analysis, Funding acquisition, Investigation, ProjectWriting – review & editing. Keren Shen: Writing – original draft. Zhuo Yu: Writing – original draft. Huizhe Lu: Writing – original draft. Te Lin: Writing – original draft. Yaozi Song: Software. Likang Luo: Data curation, Methodology. All authors reviewed the manuscript.

Data availability statement

The datasets generated and analysed during the current study are not publicly available due to hospital data security policy but are available from the corresponding author on reasonable request.

Competing interests

The author(s) declare no competing interests.

Wang W, Jiang B, Sun H, et al. Prevalence, incidence, and mortality of stroke in China, results from nationwide population-based survey of 480687 adults. Circulation, 2017, 135(8): 758-771.
Wang D, Liu J, Liu M, et al. Patterns of stroke between university hospital and nonuniversity hospitals in mainland China: prospective multicenter hospital-based registry study. World Neurosurg, 2017, 98: 258-265.
Mohamad A, James ES, Jin S. et al. Neuroimaging of Acute Ischemic Stroke: Multimodal Imaging Approach for Acute Endovascular Therapy. J Stroke. 2023 Jan; 25(1): 55–71.
Zerna C, Thomalla G, Campbell BCV, et al. Current practice and future directions in the diagnosis and acute treatment of ischaemic stroke. Lancet. 2018;392(10154):1247.
Alshoabi S, Alnajmani R, Shamsuddin M, et al. Early signs of middle cerebral artery infarction on multidetector computed tomography: Review of 20 cases[J]. Brain Circulation, 2019, 5(1): 27.
Zhang W, Cheng J, Zhang Y, et al. Analysis of CT and MRI combined examination for the diagnosis of acute cerebral infarction[J]. J Coll Physicians Surg Pak, 2019, 29(9): 898-9.
Wardlaw JM, Mielke O. Early signs of brain infarction at CT: observer reliability and outcome after thrombolytic treatment--systematic review. Radiology. 2005;235(2):444.
Hacke W, Kaste M, Fieschi C, et al. Intravenous thrombolysis with recombinant tissue plasminogen activator for acute hemispheric stroke. The European Cooperative Acute Stroke Study (ECASS). JAMA. 1995;274(13):1017.
Wardlaw JM, Dorman PJ, Lewis SC, et al. Can stroke physicians and neuroradiologists identify signs of early cerebral infarction on CT? J Neurol Neurosurg Psychiatry. 1999;67(5):651.
Ravishankar S, Ye J C, Fessler J A. Image Reconstruction: From Sparsity to Data-adaptive Methods and Machine Learning [J]. Proceedings of the IEEE Institute of Electrical and Electronics Engineers, 2020, 108(1): 86-109.
Hammernik K, Klatzer T, Kobler E, et al. Learning a variational network for reconstruction of accelerated MRI data [J]. Magnetic resonance in medicine, 2018, 79(6): 3055-3071.
Chen F, Valentina T, Itzik M, et al. Variable-Density Single-Shot Fast Spin-Echo MRI with Deep Learning Reconstruction by Using Variational Networks [J]. Radiology, 2018, 289(2): 366-373.
Wang S, Su Z, Ying L, et al. Accelerating Magnetic Resonance Imaging Via Deep Learning [J]. Proceedings IEEE International Symposium on Biomedical Imaging, 2016, 2016: 514-517.
Kim K H, Do W J, Park S H. Improving resolution of MR images with an adversarial network incorporating images with different contrast [J]. Medical physics, 2018, 45(7): 3120-3131.
Qin C, Schlemper J, Caballero J, et al. Convolutional Recurrent Neural Networks for Dynamic MR Image Reconstruction [J]. IEEE transactions on medical imaging, 2019, 38(1): 280-290.
Qin C, Schlemper J, Duan J, et al. k-t NEXT: Dynamic MR Image Reconstruction Exploiting Spatio-temporal Correlations [C]. 22nd International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2019, 11765: 505-513.
Anupama C S S, Sivaram M, Lydia E L, et al. Synergic deep learning model–based automated detection and classification of brain intracranial hemorrhage images in wearable networks[J]. Personal and Ubiquitous Computing, 2022: 1-10.
Cameron J M, Brennan P M, Antoniou G, et al. Clinical validation of a spectroscopic liquid biopsy for earlier detection of brain cancer[J]. Neuro-oncology advances, 2022, 4(1): vdac024.
Hannan S A, Hivre M, Lata M, et al. Brain damage detection using Machine learning approach[J]. International Journal of Health Sciences, Special, 2022 (7): 4910-4924.
Hu N, Zhang T, Wu Y, et al. Detecting brain lesions in suspected acute ischemic stroke with CT-based synthetic MRI using generative adversarial networks[J]. Annals of Translational Medicine, 2022, 10(2).
Zhang M J, Zhang X, Xu Y X. Analysis on value of CT and MRI clinical application in diagnosis of middle-aged patients with multiple cerebral infarction[J]. International Journal of Clinical and Experimental Medicine, 2015, 8(10): 17123.
Phaphuangwittayakul A, Guo Y, Ying F, et al. An optimal deep learning framework for multi-type hemorrhagic lesions detection and quantification in head CT images for traumatic brain injury[J]. Applied Intelligence, 2022: 1-19.
Öman O, Mäkelä T, Salli E, et al. 3D convolutional neural networks applied to CT angiography in the detection of acute ischemic stroke[J]. European radiology experimental, 2019, 3: 1-11.
van Ommen F, Dankbaar J W, Zhu G, et al. Virtual monochromatic dual-energy CT reconstructions improve detection of cerebral infarct in patients with suspicion of stroke[J]. Neuroradiology, 2021, 63: 41-49.
KXie S, Girshick R, Dollár P, et al. Aggregated residual transformations for deep neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1492-1500.
Bansal A, Sharma R, Sharma V, et al. Detecting Severity Levels of Cucumber Leaf Spot Disease using ResNext Deep Learning Model: A Digital Image Analysis Approach[C]//2023 4th International Conference for Emerging Technology (INCET). IEEE, 2023: 1-6.
Bairaboina S S R, Battula S R. Ghost-resNeXt: an effective deep learning based on mature and immature WBC classification[J]. Applied Sciences, 2023, 13(6): 4054.
Veeramreddy R, Gnanasekaran A, Babu K S. Detection of diseases in rice leaf using convolutional neural network with transfer learning based on ResNeXt[J]. International Journal of Electrical & Computer Engineering (2088-8708), 2024, 14(2).
Targ S, Almeida D, Lyman K. Resnet in resnet: Generalizing residual architectures[J]. arXiv preprint arXiv:1603.08029, 2016.
Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7132-7141.
Zhong J L, Pun C M. An end-to-end dense-inceptionnet for image copy-move forgery detection[J]. IEEE Transactions on Information Forensics and Security, 2019, 15: 2134-2146.
Luo Y, Wang Z. An improved resnet algorithm based on cbam[C]//2021 International Conference on Computer Network, Electronic and Automation (ICCNEA). IEEE, 2021: 121-125.
Zhang X, Zhou X, Lin M, et al. Shufflenet: An extremely efficient convolutional neural network for mobile devices[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 6848-6856.
Liu Z, Lin Y, Cao Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2021: 10012-10022.
Brunser A M, Hoppe A, Illanes S, et al. Accuracy of diffusion-weighted imaging in the diagnosis of stroke in patients with suspected cerebral infarct[J]. Stroke, 2013, 44(4): 1169-1171.
Gao P, Shan W, Guo Y, et al. Development and validation of a deep learning model for brain tumor diagnosis and classification using magnetic resonance imaging[J]. JAMA Network Open, 2022, 5(8): e2225608-e2225608.
Abbasi H, Orouskhani M, Asgari S, et al. Automatic brain ischemic stroke segmentation with deep learning: A review[J]. Neuroscience Informatics, 2023: 100145.
Altmann S, Grauhan N F, Brockstedt L, et al. Ultrafast brain MRI with deep learning reconstruction for suspected acute ischemic stroke[J]. Radiology, 2024, 310(2): e231938.
Verclytte S, Gnanih R, Verdun S, et al. Ultrafast MRI using deep learning echoplanar imaging for a comprehensive assessment of acute ischemic stroke[J]. European Radiology, 2023, 33(5): 3715-3725.
Kim B K, Park S, Han M K, et al. Deep learning for prediction of mechanism in acute ischemic stroke using brain diffusion magnetic resonance image[J]. Journal of Neurocritical Care, 2023, 16(2): 85-93.
Yang T H, Su Y Y, Tsai C L, et al. Magnetic resonance imaging-based deep learning imaging biomarker for predicting functional outcomes after acute ischemic stroke[J]. European journal of radiology, 2024, 174: 111405.
Alshehri F, Muhammad G. A few-shot learning-based ischemic stroke segmentation system using weighted MRI fusion[J]. Image and Vision Computing, 2023, 140: 104865.
Brugnara G, Baumgartner M, Scholze E D, et al. Deep-learning based detection of vessel occlusions on CT-angiography in patients with suspected acute ischemic stroke[J]. Nature Communications, 2023, 14(1): 4938.
Mohapatra S, Lee T H, Sahoo P K, et al. Localization of early infarction on non-contrast CT images in acute ischemic stroke with deep learning approach[J]. Scientific Reports, 2023, 13(1): 19442.

No competing interests reported.

Download PDF

Editor invited by journal
05 Sep, 2024
Submission checks completed at journal
05 Sep, 2024
First submitted to journal
06 Aug, 2024

You are reading this latest preprint version

Development of a diagnostic model for acute ischemic stroke early identification based on SE-ResNeXt

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

Introduction

Methods

Patients

Image enhancement and image preprocessing

Deep learning model development

Model evaluation

Statistical Analysis

Results

Distribution of data

Model building

Model performance evaluation

Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1