Integrating Clinical Data and Radiomics and Deep Learning Features for End-to-End Delayed Cerebral Ischemia Prediction on Noncontrast CT

Qi-qi Ban; Hao-tian Zhang; Wei Wang; Yi-fan Du; Yi Zhao; Ai-jun Peng; Hang Qu

doi:10.3174/ajnr.A8301

Abstract

BACKGROUND AND PURPOSE: Delayed cerebral ischemia is hard to diagnose early due to gradual, symptomless development. This study aimed to develop an automated model for predicting delayed cerebral ischemia following aneurysmal SAH on NCCT.

MATERIALS AND METHODS: This retrospective study included 400 patients with aneurysmal SAH (156 with delayed cerebral ischemia) who underwent NCCT. The study used ATT-Deeplabv3+ for automatically segmenting hemorrhagic regions using semisupervised learning. Principal component analysis was used for reducing the dimensionality of deep learning features extracted from the average pooling layer of ATT-DeepLabv3+. The classification model integrated clinical data, radiomics, and deep learning features to predict delayed cerebral ischemia. Feature selection involved Pearson correlation coefficients, least absolute shrinkage, and selection operator regression. We developed models based on clinical features, clinical-radiomics, and a combination of clinical, radiomics, and deep learning. The study selected logistic regression, Naive Bayes, Adaptive Boosting (AdaBoost), and multilayer perceptron as classifiers. The performance of segmentation and classification models was evaluated on their testing sets using the Dice similarity coefficient for segmentation, and the area under the receiver operating characteristic curve (AUC) and calibration curves for classification.

RESULTS: The segmentation process achieved a Dice similarity coefficient of 0.91 and the average time of 0.037 s/image. Seventeen features were selected to calculate the radiomics score. The clinical-radiomics–deep learning model with multilayer perceptron achieved the highest AUC of 0.84 (95% CI, 0.72–0.97), which outperformed the clinical-radiomics model (P = .002) and the clinical features model (P = .001) with multilayer perceptron. The performance of clinical-radiomics–deep learning model using AdaBoost was significantly superior to its clinical-radiomics model (P = .027). The performance of the clinical-radiomics–deep learning model and the clinical-radiomics model with logistic regression notably exceeded that of the model based solely on clinical features (P = .028; P = .046). The AUC of the clinical-radiomics–deep learning model with multilayer perceptron (P < .001) and the clinical-radiomics model with logistic regression (P = .046) were significantly higher than the clinical model with logistic regression. Of all models, the clinical-radiomics–deep learning model with multilayer perceptron showed best calibration.

CONCLUSIONS: The proposed 2-stage end-to-end model not only achieves rapid and accurate segmentation but also demonstrates superior diagnostic performance with high AUC values and good calibration in the clinical-radiomics–deep learning model, suggesting its potential to enhance delayed cerebral ischemia detection and treatment strategies.

ABBREVIATIONS:

AdaBoost: Adaptive Boosting
aSAH: aneurysmal subarachnoid hemorrhage
AUC: area under the receiver operating characteristic curve
DCI: delayed cerebral ischemia
DL: deep learning
DSC: Dice similarity coefficient
GCS: Glasgow Coma Scale
LR: logistic regression
ML: machine learning
MLP: multilayer perceptron
Rad: radiomics

SUMMARY

PREVIOUS LITERATURE:

Recent studies have employed Machine Learning models to predict Delayed Cerebral Ischemia (DCI), utilizing clinical predictors such as age, sex, clinical grades, aneurysm treatment, and lab tests.^1⇓–3 Image features from noncontrast CT (NCCT), have been highlighted for enhancing predictive accuracy.³ Automated segmentation techniques for assessing blood volume in subarachnoid hemorrhage (SAH) patients are also being explored, though reports on dice similarity coefficients (DSC) are often lacking or not high.^4⇓–6

KEY FINDINGS:

Image features from NCCT significantly improve DCI prediction.³ The effectiveness of automated segmentation in assessing blood volume in SAH patients is recognized, yet the need for improved reporting on segmentation accuracy is evident.^4⇓–6

KNOWLEDGE ADVANCEMENT:

This study presents end-to-end models that enhance DCI prediction by integrating clinical, radiomic, and deep learning (DL) features. With a DSC of 0.91 and a segmentation speed of 0.037s/image, it shows swift and precise segmentation. The clinical-Rad-DL model notably achieves a peak AUC of 0.84, significantly boosting DCI prediction accuracy.

Aneurysmal subarachnoid hemorrhage (aSAH) remains a destructive disease with a mortality as high as 35% and a good chance of long-term dysfunction in survivors.¹ Delayed cerebral ischemia (DCI) occurs in 20%–30% of patients with aSAH and continues to play an important role in poor outcomes due to late diagnosis.^2,3 Early detection of DCI can influence patient outcomes and decrease the expenses associated with intensive care.

The diagnosis of DCI after aSAH is challenging, primarily due to its gradual onset and frequently asymptomatic nature.⁴ Current imaging methods like invasive conventional angiography and less invasive CT scans focus on macrovascular vasospasm. However, vasospasm is not a key treatment target for the patient’s functional outcome.⁵ Other techniques like transcranial Doppler and continuous electroencephalography face limitations in accuracy and infrastructure.⁶

Recently, some machine learning (ML) models have already been used to forecast DCI.^7⇓-9 A review confirmed that some clinical and radiologic predictors are associated with DCI.¹⁰ Subsequent studies typically encompassed predictors such as age, sex, clinical grades, treatment of aneurysms, and laboratory test results for predicting DCI.^7,11,12 The incorporation of image features particularly enhances the predictive accuracy.⁹ Shan et al¹³ predicted the outcome of SAH by analyzing radiomics features derived from NCCT. Furthermore, the application of automated segmentation techniques for evaluating blood volume in patients with aSAH has been suggested.^14⇓-16 However, the Dice similarity coefficients (DSCs)¹⁷ in these studies often fell short or were unreported.

Currently, NCCT scans are the primary diagnostic tool for SAH.¹⁸ However, the traditional interpretation of NCCT images provides a relatively limited perspective of informative insights. The future direction emphasizes the integration of current methods and the development of new techniques.

Radiomics (Rad), a novel method in medical problem-solving, converts medical images into high-throughput features, extracting additional, unused information.¹⁹ Deep learning (DL) approaches adaptively learn and extract more feature information directly from a vast amount of data beyond predefined features.²⁰ Among the DL architectures, convolutional neural networks are predominantly used in computer vision,²¹ showing promising results in medical image-analysis tasks.²² The advancement of DL models in medical image analysis is significantly hindered by the scarcity of extensive and meticulously annotated data sets; to address this issue, semisupervised learning methods are increasingly used for generating more labeled data, uncovering hidden patterns in unlabeled images, and creating pseudolabels for further model optimization.^23,24

In this study, we proposed 2-stage end-to-end models that integrate clinical, radiomics, and DL features to predict DCI after aSAH. We constructed 3 distinct models: the clinical features model, the clinical-radiomics model (the clinical-Rad model), and the clinical-Rad-DL model. We hypothesized that automated segmentation will ensure fast and accurate ROI identification in NCCT images, radiomics features will improve DCI prediction performance beyond clinical data alone, and incorporating DL features will further enhance the diagnostic performance of the model.

MATERIALS AND METHODS

Study Population

Our study encompassed all adult patients admitted with SAH who underwent NCCT before any treatment or operation at the affiliated hospital of Yangzhou university from 2014 to 2022. We focused on those with Fisher grade 3 SAH, the group most at risk for developing DCI.²⁵ The other inclusion criterion was the causative aneurysm verified through CTA, MRA, DSA, or surgical intervention. Initially, 478 patients were considered for inclusion. However, 78 were excluded due to various reasons including no scan within 24 hours of ictus (n = 31), severe imaging artifacts (n = 32), missing clinical information (n = 11), and duplicate admission (n = 4). Ultimately, 400 patients were included, with 156 (39%) developing DCI. Inclusion and exclusion criteria and grouping of patients are shown in Fig 1.

FIG 1.

Inclusion and exclusion criteria and grouping of patients.

Clinical Data

We collected baseline clinical data comprising age, sex, time from onset to NCCT, history of hypertension, diabetes, smoking, aneurysms, diastolic and systolic pressure, white blood cell count, blood glucose level, systemic inflammatory response syndrome, aneurysm treatment modalities, and size and location of aneurysms. Disease and hemorrhage severity were independently evaluated by 2 trained physicians using the World Federation of Neurosurgical Societies scale, Hunt and Hess grade, Glasgow Coma Scale (GCS), and the modified Fisher grade based on clinical status and imaging findings. Discrepancies in assessments were resolved through team consensus.

Outcome Definitions

DCI is defined by fulfilling any one of the following criteria: 1) the presence of a new, lasting, or transient focal neurologic deficits (eg, aphasia, apraxia, hemianopia, or neglect) occurring between the fourth and 14th day after aSAH, without any other identifiable cause; 2) a reduction in the GCS by at least 2 points, which could be due to changes in any of its components (such as eye opening, verbal response, or motor response) or the overall score, and head CT showing a new area of low density not present at admission or immediately postsurgery, with vasospasms as the only identified cause within 4–30 days following aSAH.^3,26

Image Acquisition

Unenhanced brain CT images were acquired using a Somatom Definition Flash or Somatom Force CT machine (Siemens). All scans used automated tube voltage selection (Care kV; Siemens), with a quality reference tube voltage of 120 kV(peak). Automatic tube current modulation (CARE Dose 4D; Siemens) was applied with a quality reference tube current time product of 330 mAs (Flash) and 273 mAs (Force). Acquired images were stored in DICOM format and analyzed using ITK-SNAP software (Version 4.0.2; http://www.itksnap.org).

Construction and Training of the Segmentation Model

Network Architecture.

We enhanced the conventional Deeplabv3+ architecture²⁷ with a novel extension called ATT-Deeplabv3+ (Fig 2), which incorporates a self-attention module into the deep convolutional neural network. A detailed explanation of the ATT-Deeplabv3+ architecture can be found in the Online Supplemental Data.

FIG 2.

The overview of the ATT-DeeplabV3+. This model consists of an encoder, leveraging a DCNN with a self-attention module for feature extraction, and a decoder for precise segmentation and upsampling. The self-attention module enhances the ability of the network to capture long-range dependencies, resulting in more accurate segmentation. DCNN indicates deep convolutional neural network.

Semisupervised Learning.

Four hundred cases were randomly split into a training set of 320 cases (including 80 labeled and 240 unlabeled cases) and a testing set of 80 cases. This division allowed us to leverage a semisupervised learning approach, using both labeled and unlabeled data for training to enhance model performance, while the separate testing set ensures the evaluation of the performance of the model.

Our study uses a self-training methodology.²³ The process begins with training an initial model on the available labeled cases. This model then serves as a basis for generating pseudolabels for the unlabeled cases, effectively using the current understanding of the model to infer labels for data points without them. These pseudolabeled data points are then added to the training data set, enlarging the pool of data from which the model can learn. Through successive iterations, the model iteratively refines the pseudolabels on the basis of its evolving understanding, gradually improving segmentation accuracy. This iterative refinement continues until the performance of the model stabilizes.

Training involves resizing input images to 256 × 256 pixels, applying z score normalization, and augmenting data with random horizontal flipping and cropping. The Stochastic Gradient Descent optimizer, with an initial learning rate of 0.01 adjusted by cosine annealing over 200 epochs, updates the model parameters. Each epoch includes 1800 iterations, and the batch size is set at 128.

Our training process featured a manual review phase to ensure segmentation accuracy, with experts refining both labeled and select pseudolabeled cases, enhancing data fidelity and optimizing semisupervised learning efficiency through a mix of automation and manual corrections.

ITK-SNAP software (Version 4.0.2) was used by 2 senior radiologists to manually draw the hemorrhagic area on NCCT after consultation. Our research focused on segmenting cisternal SAH due to its significant association with the risk of DCI.¹⁶ Manual segmentation involved 80 labeled patients and another 80 patients in the testing set.

Radiomics Feature and DL Feature Extraction

The PyRadiomics open-source library (Version 3.1.0) (https://pyradiomics.readthedocs.io/) was applied to extract Rad features. This advanced tool enabled us to extract a comprehensive set of 107 Rad features from the automated segmented ROI for each patient. These features included first-order features, shape-based features, gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix, neighboring gray-tone difference matrix (NGTDM), and gray-level dependence matrix.

Principal component analysis is used for reducing the dimensionality of DL features from the average pooling layer of DeepLabv3+, compressing 2048 features per patient into 31 features. These components are the direct outcome of the ability of principal component analysis to identify and retain the most significant variance and information from the original features.²⁸

Feature Selection and Classification Model Construction

We integrated 31 DL features with clinical and radiomics features for feature selection. The patients were divided into a training set (n = 320) and a testing set (n = 80) at a ratio of 8:2 via stratified random sampling.

After standardizing the data, we computed the Pearson correlation coefficients among the features. We retained features with a correlation coefficient of >0.9 within the training cohort (Online Supplemental Data). Subsequently, the least absolute shrinkage and selection operator regression (LASSO; https://www.ibm.com/topics/lasso-regression) was applied to select the features that are most important for predicting DCI.²⁹ LASSO selects influential features by shrinking less important regression coefficients to zero with a penalty term, leading to increased mean squared error as the penalty grows. Our goal was to identify an optimal penalty level that minimizes variables while maintaining a low mean squared error.

We developed models based on clinical features (clinical features model), clinical and radiomics features (clinical-Rad model), and a combination of clinical-radiomics-DL features (clinical-Rad-DL model), respectively.

For the classification model (DCI prediction model), we chose logistic regression (LR), Naive Bayes, Adaptive Boosting (AdaBoost), and multilayer perceptron (MLP) to achieve a high and stable performance. To construct classifiers optimized with the most suitable hyperparameters, we used the comprehensive grid search strategy via the GridSearchCV function (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html), applying it to the training data set alongside a 10-fold cross-validation method.³⁰

Our study used 2-stage end-to-end models, beginning with an independent segmentation phase that influences the subsequent classification model. Most important, segmentation and classification data sets are unrelated and chosen separately. Training and testing data sets were kept separate to prevent leakage of information from the testing data set.

Statistical Analysis

R software (R Foundation for Statistical Computing) (https://www.r-project.org/) and SPSS Statistics (Version 27) (https://www.ibm.com/products/spss-statistics) were used for statistical analysis. A Student t test was used for normally distributed continuous variables, and the Mann-Whitney test was used for other continuous variables. The χ² test or Fisher exact probability method was used for count data. The performance of the segmentation model was measured by DSC, which quantifies the volume overlap between expert manual segmentation and the automatic segmentation produced by the model for the hemorrhagic regions. The performance of the classification model was assessed using the sensitivity, specificity, recall, accuracy, and area under the receiver operating characteristic curve (AUC). Differences in the AUC values between different models were estimated using the Delong test. P < .05 was considered statistically significant.

Compliance With Ethical Standards

Institutional review board approval was obtained by ethics committee of affiliated hospital of Yangzhou university (Approval Number:2022-YKL8-07).

RESULTS

Baseline Characteristics

Clinical and imaging indices are shown in the Online Supplemental Data. Statistically significant differences between the DCI and non-DCI groups were observed in age, history of smoking, white blood cell count, blood glucose levels, GCS, World Federation of Neurosurgical Societies scale, Hunt and Hess grade, systemic inflammatory response syndrome, and modified Fisher scores. For the other characteristics, there were no statistical differences between DCI and non-DCI groups. No statistical differences were observed in clinical and imaging indices between the training and testing groups.

Performance of Segmentation

The average DSC was obtained as 0.91, and the average time to segment is 0.037s/image.

As shown in Fig 3, our model performed well by segmenting the regions similar to the ground truth. Especially in cases 1 and 4, our model successfully managed to eliminate the interference of bone tissue in the identification of the ROI.

FIG 3.

Visualization of segmentation results predicted by ATT-Deeplabv3+ model. Cases 1–4 are shown, arranged from upper to lower rows. Column 1 represents the original image, column 2 represents the ground truth, and column 2 represents the results of the ATT-Deeplabv3+ model.

In addition, we compared our segmentation model with previously reported models (Online Supplemental Data).

Construction of the Radiomics Scoring Model

Seventeen features with nonzero coefficients were ultimately selected using the LASSO regression model (λ = 0.0596) (Online Supplemental Data). The details about the features attributed to the radiomics score (Rad-score) are summarized in the Online Supplemental Data.

The coefficient values for the selected features are shown in Fig 4.

FIG 4.

Coefficient values for the selected features. The y-axis indicates the selected 17 features, and the x-axis represents the coefficient of features. Smoke indicates a history of smoking; SIRS, systemic inflammatory response syndrome; Glu, glucose; HP, history of hypertension; WFANS, World Federation of Neurosurgical Societies.

Performance of Models

Among the 4 ML classifiers, LR yielded the highest AUC of 0.66 (95% CI, 0.48–0.84) in the testing set of the clinical features model, with a sensitivity and recall of 1.00. In the testing set of the clinical-Rad model, LR again showed the highest AUC of 0.77 (95% CI, 0.62–0.92), with the accuracy of 0.73, sensitivity of 0.77, specificity of 0.70, and recall of 0.77. In the training set of the clinical-Rad-DL model, the MLP yielded an AUC of 0.87 (95% CI, 0.82–0.92). Furthermore, in the testing set of the clinical-Rad-DL model, the MLP achieved the highest AUC of 0.84 (95% CI, 0.72–0.97), with a sensitivity and recall of 1.00 and an accuracy of 0.70. Details of each model are shown in Fig 5 (Online Supplemental Data).

FIG 5.

The performance of all models. A, ROC curves of the clinical features model. B, ROC curves of the clinical-radiomics model. C, ROC curves of the clinical, radiomics, DL model.

The calibration curve showed good calibration in the testing sets of the clinical-rad-DL model (Fig 6). The calibration curves of other models are shown in the Online Supplemental Data.

FIG 6.

The calibration curves of clinical-radiomics-DL models in the testing set. The x-axis represents the outcome predicted by the models, and the y-axis shows the actual outcome. The 45° straight line represents the ideal evaluation by a perfect model. The dotted line illustrates the performance of the model, and the solid line corrects for any bias in the model. The closer the line of the model is to the 45°straight line, the better is its evaluation. A, The calibration curves of clinical-radiomics-DL models with LR. B, The calibration curves of clinical-radiomics-DL models with Naive Bayes. C, The calibration curves of clinical, radiomics, DL models with AdaBoost. D, The calibration curves of clinical-radiomics-DL models with MLP.

Comparison of Performance among Different Models

In the testing set of LR, the performance of the clinical-Rad model and the clinical-Rad-DL model was significantly higher than that of the clinical features model (0.77 versus 0.66, Delong: P = .046; 0.81 versus 0.66, Delong: P = .028). In the testing set of AdaBoost, the performance of the clinical-Rad-DL model was significantly higher than that of the clinical-Rad model (0.72 versus 0.54, Delong: P = .027). Similarly, in the testing set of MLP, the performance of the clinical-Rad-DL model was significantly higher than that of the clinical-Rad model (0.84 versus 0.64, Delong: P = .002) and clinical features model (0.84 versus 0.63, Delong: P = .001). However, there were no significant differences in diagnostic performance in the testing set of Naive Bayes (Online Supplemental Data).

Furthermore, in the testing set of the clinical-Rad model, LR was significantly higher than in AdaBoost (0.77 versus 0.54, Delong: P = .011). In the testing set of the clinical features model, there were no significant differences in the diagnostic performance among the 4 ML classifiers. Similarly, no significant differences were observed in the testing set of clinical-Rad-DL model.

Regarding comparison between the best-performing models from different input in test cohort, the AUC of the clinical-Rad-DL model with MLP (0.84 versus 0.66, Delong: P < .001) and the clinical-Rad model with LR (0.77 versus 0.66, Delong: P = .046) were significantly higher than the clinical model with LR. There were no significant differences in diagnostic performance between the clinical-Rad model with LR and the clinical-Rad-DL model with MLP (Online Supplemental Data).

DISCUSSION

This study introduces 2-stage end-to-end models that improve the DCI prediction after aSAH by combining clinical, radiomics, and DL features. Achieving a DSC of 0.91 and segmentation speed of 0.037 s/image, the study demonstrates rapid and accurate hemorrhagic region segmentation. Incorporating radiomics and DL features notably increases the DCI prediction accuracy, with the clinical-Rad-DL model achieving the highest testing set AUC of 0.84.

In comparison with previous studies,^14⇓-16 our segmentation results highlight the ability to perform rapid and precise segmentation, even with limited training data. Bai et al²³ demonstrated the effectiveness of the semisupervised method in segmentation tasks, particularly with limited data. In our study, we used the DeepLabv3+ model for segmentation primarily due to its proved efficiency and adaptability in handling medical imaging data.^31⇓-33

Our study uses hemorrhagic components to predict DCI, motivated by research showing a correlation between hemorrhage severity and distribution after aSAH and DCI risk.^16,34,35 Boers et al¹⁵ demonstrated the value of assessing blood volume and density in NCCT scans for improving patient outcome predictions. Shan et al¹³ demonstrated the efficacy of extracting radiomics features from NCCT images to predict the prognosis of patients with SAH.

Our approach uniquely incorporates radiomics features into the DCI prediction model. Earlier research had predominantly concentrated on clinical data and characteristics observed in medical imaging. Ramos et al⁹ integrated clinical data with image features on the basis of unsupervised learning, achieving an AUC of 0.74 in predicting DCI. Compared with the previous study, we included a larger case cohort and achieved a higher AUC. Furthermore, the incorporation of radiomics in our model enhances its interpretability.

After we added DL features, models showed statistically better performance, which demonstrates that these DL features can capture unique information beyond the scope of manually crafted features.³⁶

Our results also highlighted factors like age, blood glucose levels, history of hypertension, GCS, Hunt and Hess grade, and the World Federation of Neurosurgical Societies scale being associated with DCI occurrence, aligning with prior studies.^{7,10⇓-12,26,37} Our findings underscore the significance of smoking history as a pivotal predictor of DCI. A meta-analysis has established smoking as a persistent risk factor for DCI in patients with ruptured aneurysms.³⁸ Acampa’s study showed a higher stiffness index in patients with DCI as an independent prognosticator for SAH-associated DCI,³⁹ complemented by Doonan et al,⁴⁰ who suggested that nicotine escalates arterial wall stiffness. Dhar et al⁴¹ corroborated the severity of the systemic inflammatory response syndrome being highly correlated with the risk of delayed ischemic neurologic deficits. Unlike some previous prediction models,^7,26 our ML models did not find white blood cell count as the most salient parameter, possibly due to the more comprehensive criteria of systemic inflammatory response syndrome offering richer information. Regarding age, systematic reviews reported a paradoxical association with DCI.¹⁰ Abdel-Tawab et al⁴² found that age was not significant for DCI risk. Notably, in our study, the DCI group was statistically younger than the non-DCI group. These conflicting findings might be attributed to variations in the demographic characteristics of the patients.

Focusing exclusively on patients with Fisher grade 3 in our study may have somewhat diminished the impact of total blood volume in predicting DCI, explaining why modified Fisher scores and radiomics features such as shape_VoxelVolume did not appear as final predictors in our predictive model. Research by Friedman et al³⁵ has shown that, even when limited to Fisher grade 3, ventricular blood volume remains a key predictor of a delayed ischemic neurologic deficit, differing from our findings. Our predictive model, which includes a broader set of factors, did not identify modified Fisher scores and radiomics features such as shape_VoxelVolume as final predictors. This discrepancy might be due to the ambiguities in the modified Fisher scale’s definitions of “thin” and “thick” SAH, which lead to variability in interpretations among raters. Woo et al⁴³ reported only moderate interobserver agreement for the modified system. Van der Steen et al⁴⁴ found that total blood volume discriminated better than the modified Fisher scale but still was only moderately predictive of the development of DCI. Our findings suggest the need for more objective and quantifiable imaging markers in the development of robust predictive models for DCI.

We identified the original_shape_Maximum3DDiameter as the most critical radiomics predictor, which indicates the maximum size or diameter of a 3D object or structure in the medical image. Previous studies have demonstrated a significant relationship between the quantified total blood volume and the likelihood of DCI development.^35,45 This predictor suggests not only the size but also the shape of the hemorrhagic region. Additionally, the original_ngtdm_Complexity was identified as a key negative predictor, indicating that less complex regions, with uniform gray-level intensity, are more prone to acquire DCI. Shan et al¹³ also found a negative correlation with ngtdm_busyness, a similar NGTDM feature, in their ML-based prognostic model for SAH. Our findings underscore the potential of NCCT-based radiomics analysis, offering a novel avenue for risk stratification.

Although the performance of the current model is not yet adequate for clinical deployment, it demonstrates the potential of artificial intelligence in revealing hidden information in NCCT images. Enhancements to artificial intelligence–based prediction models can be achieved using additional information from the individually analyzed blood volumes of brain compartments and by considering the surrounding cerebral parenchymal edema.

Our study has some limitations. The retrospective design and single-center approach of the study might introduce biases, and the number of participants could limit the broader applicability of the findings. To validate our conclusions, prospective studies involving larger populations across multiple centers are essential. The study included only patients with Fisher grade 3, with most cases having hemorrhage limited to the subarachnoid space. This limitation may introduce a certain selection bias. In the future, selecting a larger sample that includes all Fisher grades and extracting imaging features from different hemorrhagic regions to provide a wider range of potential characteristics should be considered.

CONCLUSIONS

In our study, we proposed a 2-stage end-to-end CT-based model for predicting DCI after aSAH with rapid and accurate segmentation of the hemorrhagic region. Incorporating radiomic features elevates the precision of DCI prediction beyond the capabilities of stand-alone clinical models. The introduction of DL features further optimizes the diagnostic efficacy of the model. The Clinical-Rad-DL model showed great diagnostic performance with a high AUC value and good calibration. The results indicated that our model may assist clinicians in detecting DCI, which is important for optimizing the overall management of patients with aSAH.

Footnotes

↵# Qi-qi Ban, Hao-tian Zhang, and Wei Wang contributed equally to this work.
This work was funded by the Program of Jiangsu Commission of Health (No. M2022068) and the Social Development Project of Yangzhou Science and Technology Bureau (No. YZ2022078).
Disclosure forms provided by the authors are available with the full text and PDF of this article at www.ajnr.org.

References

1.↵
1. Neifert SN,
2. Chapman EK,
3. Martini ML, et al
. Aneurysmal subarachnoid hemorrhage: the last decade. Transl Stroke Res 2021;12:428–46 doi:10.1007/s12975-020-00867-0 pmid:33078345
CrossRef PubMed
2.↵
1. Galea JP,
2. Dulhanty L,
3. Patel HC
, UK and Ireland Subarachnoid Hemorrhage Database Collaborators. Predictors of outcome in aneurysmal subarachnoid hemorrhage patients: observations from a multicenter data set. Stroke 2017;48:2958–63 doi:10.1161/STROKEAHA.117.017777 pmid:28974630
Abstract/FREE Full Text
3.↵
1. Vergouwen MDI,
2. Vermeulen M,
3. van Gijn J, et al
. Definition of delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage as an outcome event in clinical trials and observational studies: proposal of a multidisciplinary research group. Stroke 2010;41:2391–95 doi:10.1161/STROKEAHA.110.589275 pmid:20798370
Abstract/FREE Full Text
4.↵
1. Schmidt JM,
2. Wartenberg KE,
3. Fernandez A, et al
. Frequency and clinical impact of asymptomatic cerebral infarction due to vasospasm after subarachnoid hemorrhage. J Neurosurg 2008;109:1052–59 doi:10.3171/JNS.2008.109.12.1052 pmid:19035719
CrossRef PubMed
5.↵
1. Etminan N,
2. Vergouwen MD,
3. Ilodigwe D, et al
. Effect of pharmaceutical treatment on vasospasm, delayed cerebral ischemia, and clinical outcome in patients with aneurysmal subarachnoid hemorrhage: a systematic review and meta-analysis. J Cereb Blood Flow Metab 2011;31:1443–51 doi:10.1038/jcbfm.2011.7 pmid:21285966
CrossRef PubMed
6.↵
1. Abdulazim A,
2. Heilig M,
3. Rinkel G, et al
. Diagnosis of delayed cerebral ischemia in patients with aneurysmal subarachnoid hemorrhage and triggers for intervention. Neurocrit Care 2023;39:311–19 doi:10.1007/s12028-023-01812-3 pmid:37537496
CrossRef PubMed
7.↵
1. Savarraj JPJ,
2. Hergenroeder GW,
3. Zhu L, et al
. Machine learning to predict delayed cerebral ischemia and outcomes in subarachnoid hemorrhage. Neurology 2021;96:e553–62 doi:10.1212/WNL.0000000000011211 pmid:33184232
Abstract/FREE Full Text
8.↵
1. de Jong G,
2. Aquarius R,
3. Sanaan B, et al
. Prediction models in aneurysmal subarachnoid hemorrhage: forecasting clinical outcome with artificial intelligence. Neurosurgery 2021;88:E427–34 doi:10.1093/neuros/nyaa581 pmid:33548918
CrossRef PubMed
9.↵
1. Ramos LA,
2. van der Steen WE,
3. Sales Barros R, et al
. Machine learning improves prediction of delayed cerebral ischemia in patients with subarachnoid hemorrhage. J Neurointerv Surg 2019;11:497–502 doi:10.1136/neurintsurg-2018-014258 pmid:30415227
Abstract/FREE Full Text
10.↵
1. de Rooij NK,
2. Rinkel GJ,
3. Dankbaar JW, et al
. Delayed cerebral ischemia after subarachnoid hemorrhage: a systematic review of clinical, laboratory, and radiological predictors. Stroke 2013;44:43–54 doi:10.1161/STROKEAHA.112.674291 pmid:23250997
Abstract/FREE Full Text
11.↵
1. Chen HY,
2. Elmer J,
3. Zafar SF, et al
. Combining transcranial Doppler and EEG data to predict delayed cerebral ischemia after subarachnoid hemorrhage. Neurology 2022;98:e459–69 doi:10.1212/WNL.0000000000013126 pmid:34845057
Abstract/FREE Full Text
12.↵
1. Megjhani M,
2. Terilli K,
3. Weiss M, et al
. Dynamic detection of delayed cerebral ischemia: a study in 3 centers. Stroke 2021;52:1370–79 doi:10.1161/STROKEAHA.120.032546 pmid:33596676
CrossRef PubMed
13.↵
1. Shan D,
2. Wang J,
3. Qi P, et al
. Non-contrasted CT radiomics for SAH prognosis prediction. Bioengineering (Basel) 2023;10:967 doi:10.3390/bioengineering10080967 pmid:37627852
CrossRef PubMed
14.↵
1. Yuan JY,
2. Chen Y,
3. Jayaraman K, et al
. Automated quantification of compartmental blood volumes enables prediction of delayed cerebral ischemia and outcomes after aneurysmal subarachnoid hemorrhage. World Neurosurg 2023;170:e214–22 doi:10.1016/j.wneu.2022.10.105 pmid:36323345
CrossRef PubMed
15.↵
1. Boers AM,
2. Zijlstra IA,
3. Gathier CS, et al
. Automatic quantification of subarachnoid hemorrhage on noncontrast CT. AJNR Am J Neuroradiol 2014;35:2279–86 doi:10.3174/ajnr.A4042 pmid:25104292
Abstract/FREE Full Text
16.↵
1. van der Steen WE,
2. Zijlstra IA,
3. Verbaan D, et al
. Association of quantified location-specific blood volumes with delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage. AJNR Am J Neuroradiol 2018;39:1059–64 doi:10.3174/ajnr.A5626 pmid:29650786
Abstract/FREE Full Text
17.↵
1. Zou KH,
2. Warfield SK,
3. Bharatha A, et al
. Statistical validation of image segmentation quality based on a spatial overlap index: scientific reports. Acad Radiol 2004;11:178–89 doi:10.1016/s1076-6332(03)00671-8 pmid:14974593
CrossRef PubMed
18.↵
1. Perry JJ,
2. Sivilotti ML,
3. Émond M, et al
. Prospective implementation of the Ottawa Subarachnoid Hemorrhage Rule and 6-Hour Computed Tomography Rule. Stroke 2020;51:424–30 doi:10.1161/STROKEAHA.119.026969 pmid:31805846
CrossRef PubMed
19.↵
1. Gillies RJ,
2. Kinahan PE,
3. Hricak H
. Radiomics: images are more than pictures, they are data. Radiology 2016;278:563–77 doi:10.1148/radiol.2015151169 pmid:26579733
CrossRef PubMed
20.↵
1. Scapicchio C,
2. Gabelloni M,
3. Barucci A, et al
. A deep look into radiomics. Radiol Med 2021;126:1296–311 doi:10.1007/s11547-021-01389-x pmid:34213702
CrossRef PubMed
21.↵
1. Krizhevsky A,
2. Sutskever I,
3. Hinton G
. ImageNet classification with deep convolutional neural networks. Commun ACM 2017;60(6):84–90 doi:10.1145/3065386
CrossRef
22.↵
1. Turhon M,
2. Li M,
3. Kang H, et al
. Development and validation of a deep learning model for prediction of intracranial aneurysm rupture risk based on multi-omics factor. Eur Radiol 2023;33:6759–70 doi:10.1007/s00330-023-09672-3 pmid:37099175
CrossRef PubMed
23.↵
1. Bai W,
2. Oktay O,
3. Sinclair M, et al
. Semi-supervised learning for network-based cardiac MR image segmentation. In: Proceedings of the Medical Image Computing and Computer-Assisted Intervention−MICCAI 2017: 20th International Conference. September 11–13, 2017. Quebec City, Canada
24.↵
1. Chen X,
2. Wang X,
3. Zhang K, et al
. Recent advances and clinical applications of deep learning in medical image analysis. Med Image Anal 2022;79:102444 doi:10.1016/j.media.2022.102444 pmid:35472844
CrossRef PubMed
25.↵
1. van der Steen WE,
2. Leemans EL,
3. van den Berg R, et al
. Radiological scales predicting delayed cerebral ischemia in subarachnoid hemorrhage: systematic review and meta-analysis. Neuroradiology 2019;61:247–56 doi:10.1007/s00234-019-02161-9 pmid:30693409
CrossRef PubMed
26.↵
1. Hu P,
2. Li Y,
3. Liu Y, et al
. Comparison of conventional logistic regression and machine learning methods for predicting delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage: a multicentric observational cohort study. Front Aging Neurosci 2022;14:857521 doi:10.3389/fnagi.2022.857521 pmid:35783143
CrossRef PubMed
27.↵
1. Chen LC,
2. Zhu Y,
3. Papandreou G, et al
. Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). September 8–14, 2018. Munich, Germany
28.↵
1. Ma J,
2. Yuan Y
. Dimension reduction of image deep feature using PCA. J Visual Commun Image Represent 2019;63:102578 doi:10.1016/j.jvcir.2019.102578
CrossRef
29.↵
1. Lan Y,
2. Han B,
3. Zhai T, et al
. Clinical application of machine learning-based pathomics signature of gastric atrophy. Front Oncol 2024;14:1289265 doi:10.3389/fonc.2024.1289265 pmid:38476364
CrossRef PubMed
30.↵
1. Zhu H,
2. Hu M,
3. Ma Y, et al
. Multi-center evaluation of machine learning-based radiomic model in predicting disease free survival and adjuvant chemotherapy benefit in stage II colorectal cancer patients. Cancer Imaging 2023;23:74 doi:10.1186/s40644-023-00588-1 pmid:37537659
CrossRef PubMed
31.↵
1. Murugappan M,
2. Bourisly AK,
3. Prakash NB, et al
. Automated semantic lung segmentation in chest CT images using deep neural network. Neural Comput Appl 2023;35:15343–64 doi:10.1007/s00521-023-08407-1 pmid:37273912
CrossRef PubMed
32.↵
1. Du H,
2. Ling L,
3. Yu W, et al
. Convolutional networks for the segmentation of intravascular ultrasound images: evaluation on a multicenter dataset. Comput Methods Programs Biomed 2022;215:106599 doi:10.1016/j.cmpb.2021.106599 pmid:34974233
CrossRef PubMed
33.↵
1. Lin YC,
2. Lin G,
3. Pandey S, et al
. Fully automated segmentation and radiomics feature extraction of hypopharyngeal cancer on MRI using deep learning. Eur Radiol 2023;33:6548–56 doi:10.1007/s00330-023-09827-2 pmid:37338554
CrossRef PubMed
34.↵
1. de Rooij NK,
2. Greving JP,
3. Rinkel GJ, et al
. Early prediction of delayed cerebral ischemia after subarachnoid hemorrhage: development and validation of a practical risk chart. Stroke 2013;44:1288–94 doi:10.1161/STROKEAHA.113.001125 pmid:23512975
Abstract/FREE Full Text
35.↵
1. Friedman JA,
2. Goerss SJ,
3. Meyer FB, et al
. Volumetric quantification of Fisher Grade 3 aneurysmal subarachnoid hemorrhage: a novel method to predict symptomatic vasospasm on admission computerized tomography scans. J Neurosurg 2002;97:401–07 doi:10.3171/jns.2002.97.2.0401 pmid:12186469
CrossRef PubMed
36.↵
1. Huang EP,
2. O’Connor JP,
3. McShane LM, et al
. Criteria for the translation of radiomics into clinically useful tests. Nat Rev Clin Oncol 2023;20:69–82 doi:10.1038/s41571-022-00707-0 pmid:36443594
CrossRef PubMed
37.↵
1. Roederer A,
2. Holmes JH,
3. Smith MJ, et al
. Prediction of significant vasospasm in aneurysmal subarachnoid hemorrhage using automated data. Neurocrit Care 2014;21:444–50 doi:10.1007/s12028-014-9976-9 pmid:24715326
CrossRef PubMed
38.↵
1. Jia C,
2. Yang Y,
3. Xia Y, et al
. Scaling up visual and vision-language representation learning with noisy text supervision. In: Proceedings of the International Conference on Machine Learnin. July 18–24, 2021; Virtual
39.↵
1. Acampa M,
2. Bongiorno M,
3. Lazzerini PE, et al
. Increased arterial stiffness is a predictor of delayed ischaemic stroke after subarachnoid haemorrhage. Heart Lung Circ 2021;30:525–30 doi:10.1016/j.hlc.2020.07.016 pmid:32896483
CrossRef PubMed
40.↵
1. Doonan RJ,
2. Hausvater A,
3. Scallan C, et al
. The effect of smoking on arterial stiffness. Hypertens Res 2010;33:398–410 doi:10.1038/hr.2010.25 pmid:20379189
CrossRef PubMed
41.↵
1. Dhar R,
2. Diringer MN
. The burden of the systemic inflammatory response predicts vasospasm and outcome after subarachnoid hemorrhage. Neurocrit Care 2008;8:404–12 doi:10.1007/s12028-008-9054-2 pmid:18196475
CrossRef PubMed
42.↵
1. Abdel-Tawab M,
2. Hasan AA,
3. Ahmed MA, et al
. Prognostic factors of delayed cerebral ischemia after subarachnoid hemorrhage including CT perfusion: a prospective cohort study. Egypt J Radiol Nucl Med 2020;51:61 doi:10.1186/s43055-020-00180-8
CrossRef
43.↵
1. Woo PY,
2. Tse TP,
3. Chan RS, et al
. Computed tomography interobserver agreement in the assessment of aneurysmal subarachnoid hemorrhage and predictors for clinical outcome. J Neurointerv Surg 2017;9:1118–24 doi:10.1136/neurintsurg-2016-012576 pmid:29030464
Abstract/FREE Full Text
44.↵
1. van der Steen WE,
2. Marquering HA,
3. Boers AM, et al
. Predicting delayed cerebral ischemia with quantified aneurysmal subarachnoid blood volume. World Neurosurg 2019;130:e613–19 doi:10.1016/j.wneu.2019.06.170 pmid:31260850
CrossRef PubMed
45.↵
1. Zijlstra IA,
2. Gathier CS,
3. Boers AM, et al
. Association of automatically quantified total blood volume after aneurysmal subarachnoid hemorrhage with delayed cerebral ischemia. AJNR Am J Neuroradiol 2016;37:1588–93 doi:10.3174/ajnr.A4771 pmid:27102313
Abstract/FREE Full Text

Received January 30, 2024.
Accepted after revision April 3, 2024.

In this issue

Download PDF

Email Article

Citation Tools

Purchase

Cited By...

No citing articles found.

Artificial Intelligence

Show more Artificial Intelligence

Neurovascular/Stroke Imaging

Show more Neurovascular/Stroke Imaging

[1] 1.↵
Neifert SN,
Chapman EK,
Martini ML, et al
. Aneurysmal subarachnoid hemorrhage: the last decade. Transl Stroke Res 2021;12:428–46 doi:10.1007/s12975-020-00867-0 pmid:33078345
CrossRef PubMed

[2] Neifert SN,

[3] Chapman EK,

[4] Martini ML, et al

[5] 2.↵
Galea JP,
Dulhanty L,
Patel HC
, UK and Ireland Subarachnoid Hemorrhage Database Collaborators. Predictors of outcome in aneurysmal subarachnoid hemorrhage patients: observations from a multicenter data set. Stroke 2017;48:2958–63 doi:10.1161/STROKEAHA.117.017777 pmid:28974630
Abstract/FREE Full Text

[6] Galea JP,

[7] Dulhanty L,

[8] Patel HC

[9] 3.↵
Vergouwen MDI,
Vermeulen M,
van Gijn J, et al
. Definition of delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage as an outcome event in clinical trials and observational studies: proposal of a multidisciplinary research group. Stroke 2010;41:2391–95 doi:10.1161/STROKEAHA.110.589275 pmid:20798370
Abstract/FREE Full Text

[10] Vergouwen MDI,

[11] Vermeulen M,

[12] van Gijn J, et al

[13] 4.↵
Schmidt JM,
Wartenberg KE,
Fernandez A, et al
. Frequency and clinical impact of asymptomatic cerebral infarction due to vasospasm after subarachnoid hemorrhage. J Neurosurg 2008;109:1052–59 doi:10.3171/JNS.2008.109.12.1052 pmid:19035719
CrossRef PubMed

[14] Schmidt JM,

[15] Wartenberg KE,

[16] Fernandez A, et al

[17] 5.↵
Etminan N,
Vergouwen MD,
Ilodigwe D, et al
. Effect of pharmaceutical treatment on vasospasm, delayed cerebral ischemia, and clinical outcome in patients with aneurysmal subarachnoid hemorrhage: a systematic review and meta-analysis. J Cereb Blood Flow Metab 2011;31:1443–51 doi:10.1038/jcbfm.2011.7 pmid:21285966
CrossRef PubMed

[18] Etminan N,

[19] Vergouwen MD,

[20] Ilodigwe D, et al

[21] 6.↵
Abdulazim A,
Heilig M,
Rinkel G, et al
. Diagnosis of delayed cerebral ischemia in patients with aneurysmal subarachnoid hemorrhage and triggers for intervention. Neurocrit Care 2023;39:311–19 doi:10.1007/s12028-023-01812-3 pmid:37537496
CrossRef PubMed

[22] Abdulazim A,

[23] Heilig M,

[24] Rinkel G, et al

[25] 7.↵
Savarraj JPJ,
Hergenroeder GW,
Zhu L, et al
. Machine learning to predict delayed cerebral ischemia and outcomes in subarachnoid hemorrhage. Neurology 2021;96:e553–62 doi:10.1212/WNL.0000000000011211 pmid:33184232
Abstract/FREE Full Text

[26] Savarraj JPJ,

[27] Hergenroeder GW,

[28] Zhu L, et al

[29] 8.↵
de Jong G,
Aquarius R,
Sanaan B, et al
. Prediction models in aneurysmal subarachnoid hemorrhage: forecasting clinical outcome with artificial intelligence. Neurosurgery 2021;88:E427–34 doi:10.1093/neuros/nyaa581 pmid:33548918
CrossRef PubMed

[30] de Jong G,

[31] Aquarius R,

[32] Sanaan B, et al

[33] 9.↵
Ramos LA,
van der Steen WE,
Sales Barros R, et al
. Machine learning improves prediction of delayed cerebral ischemia in patients with subarachnoid hemorrhage. J Neurointerv Surg 2019;11:497–502 doi:10.1136/neurintsurg-2018-014258 pmid:30415227
Abstract/FREE Full Text

[34] Ramos LA,

[35] van der Steen WE,

[36] Sales Barros R, et al

[37] 10.↵
de Rooij NK,
Rinkel GJ,
Dankbaar JW, et al
. Delayed cerebral ischemia after subarachnoid hemorrhage: a systematic review of clinical, laboratory, and radiological predictors. Stroke 2013;44:43–54 doi:10.1161/STROKEAHA.112.674291 pmid:23250997
Abstract/FREE Full Text

[38] de Rooij NK,

[39] Rinkel GJ,

[40] Dankbaar JW, et al

[41] 11.↵
Chen HY,
Elmer J,
Zafar SF, et al
. Combining transcranial Doppler and EEG data to predict delayed cerebral ischemia after subarachnoid hemorrhage. Neurology 2022;98:e459–69 doi:10.1212/WNL.0000000000013126 pmid:34845057
Abstract/FREE Full Text

[42] Chen HY,

[43] Elmer J,

[44] Zafar SF, et al

[45] 12.↵
Megjhani M,
Terilli K,
Weiss M, et al
. Dynamic detection of delayed cerebral ischemia: a study in 3 centers. Stroke 2021;52:1370–79 doi:10.1161/STROKEAHA.120.032546 pmid:33596676
CrossRef PubMed

[46] Megjhani M,

[47] Terilli K,

[48] Weiss M, et al

[49] 13.↵
Shan D,
Wang J,
Qi P, et al
. Non-contrasted CT radiomics for SAH prognosis prediction. Bioengineering (Basel) 2023;10:967 doi:10.3390/bioengineering10080967 pmid:37627852
CrossRef PubMed

[50] Shan D,

[51] Wang J,

[52] Qi P, et al

[53] 14.↵
Yuan JY,
Chen Y,
Jayaraman K, et al
. Automated quantification of compartmental blood volumes enables prediction of delayed cerebral ischemia and outcomes after aneurysmal subarachnoid hemorrhage. World Neurosurg 2023;170:e214–22 doi:10.1016/j.wneu.2022.10.105 pmid:36323345
CrossRef PubMed

[54] Yuan JY,

[55] Chen Y,

[56] Jayaraman K, et al

[57] 15.↵
Boers AM,
Zijlstra IA,
Gathier CS, et al
. Automatic quantification of subarachnoid hemorrhage on noncontrast CT. AJNR Am J Neuroradiol 2014;35:2279–86 doi:10.3174/ajnr.A4042 pmid:25104292
Abstract/FREE Full Text

[58] Boers AM,

[59] Zijlstra IA,

[60] Gathier CS, et al

[61] 16.↵
van der Steen WE,
Zijlstra IA,
Verbaan D, et al
. Association of quantified location-specific blood volumes with delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage. AJNR Am J Neuroradiol 2018;39:1059–64 doi:10.3174/ajnr.A5626 pmid:29650786
Abstract/FREE Full Text

[62] van der Steen WE,

[63] Zijlstra IA,

[64] Verbaan D, et al

[65] 17.↵
Zou KH,
Warfield SK,
Bharatha A, et al
. Statistical validation of image segmentation quality based on a spatial overlap index: scientific reports. Acad Radiol 2004;11:178–89 doi:10.1016/s1076-6332(03)00671-8 pmid:14974593
CrossRef PubMed

[66] Zou KH,

[67] Warfield SK,

[68] Bharatha A, et al

[69] 18.↵
Perry JJ,
Sivilotti ML,
Émond M, et al
. Prospective implementation of the Ottawa Subarachnoid Hemorrhage Rule and 6-Hour Computed Tomography Rule. Stroke 2020;51:424–30 doi:10.1161/STROKEAHA.119.026969 pmid:31805846
CrossRef PubMed

[70] Perry JJ,

[71] Sivilotti ML,

[72] Émond M, et al

[73] 19.↵
Gillies RJ,
Kinahan PE,
Hricak H
. Radiomics: images are more than pictures, they are data. Radiology 2016;278:563–77 doi:10.1148/radiol.2015151169 pmid:26579733
CrossRef PubMed

[74] Gillies RJ,

[75] Kinahan PE,

[76] Hricak H

[77] 20.↵
Scapicchio C,
Gabelloni M,
Barucci A, et al
. A deep look into radiomics. Radiol Med 2021;126:1296–311 doi:10.1007/s11547-021-01389-x pmid:34213702
CrossRef PubMed

[78] Scapicchio C,

[79] Gabelloni M,

[80] Barucci A, et al

[81] 21.↵
Krizhevsky A,
Sutskever I,
Hinton G
. ImageNet classification with deep convolutional neural networks. Commun ACM 2017;60(6):84–90 doi:10.1145/3065386
CrossRef

[82] Krizhevsky A,

[83] Sutskever I,

[84] Hinton G

[85] 22.↵
Turhon M,
Li M,
Kang H, et al
. Development and validation of a deep learning model for prediction of intracranial aneurysm rupture risk based on multi-omics factor. Eur Radiol 2023;33:6759–70 doi:10.1007/s00330-023-09672-3 pmid:37099175
CrossRef PubMed

[86] Turhon M,

[87] Li M,

[88] Kang H, et al

[89] 23.↵
Bai W,
Oktay O,
Sinclair M, et al
. Semi-supervised learning for network-based cardiac MR image segmentation. In: Proceedings of the Medical Image Computing and Computer-Assisted Intervention−MICCAI 2017: 20th International Conference. September 11–13, 2017. Quebec City, Canada

[90] Bai W,

[91] Oktay O,

[92] Sinclair M, et al

[93] 24.↵
Chen X,
Wang X,
Zhang K, et al
. Recent advances and clinical applications of deep learning in medical image analysis. Med Image Anal 2022;79:102444 doi:10.1016/j.media.2022.102444 pmid:35472844
CrossRef PubMed

[94] Chen X,

[95] Wang X,

[96] Zhang K, et al

[97] 25.↵
van der Steen WE,
Leemans EL,
van den Berg R, et al
. Radiological scales predicting delayed cerebral ischemia in subarachnoid hemorrhage: systematic review and meta-analysis. Neuroradiology 2019;61:247–56 doi:10.1007/s00234-019-02161-9 pmid:30693409
CrossRef PubMed

[98] van der Steen WE,

[99] Leemans EL,

[100] van den Berg R, et al

[101] 26.↵
Hu P,
Li Y,
Liu Y, et al
. Comparison of conventional logistic regression and machine learning methods for predicting delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage: a multicentric observational cohort study. Front Aging Neurosci 2022;14:857521 doi:10.3389/fnagi.2022.857521 pmid:35783143
CrossRef PubMed

[102] Hu P,

[103] Li Y,

[104] Liu Y, et al

[105] 27.↵
Chen LC,
Zhu Y,
Papandreou G, et al
. Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). September 8–14, 2018. Munich, Germany

[106] Chen LC,

[107] Zhu Y,

[108] Papandreou G, et al

[109] 28.↵
Ma J,
Yuan Y
. Dimension reduction of image deep feature using PCA. J Visual Commun Image Represent 2019;63:102578 doi:10.1016/j.jvcir.2019.102578
CrossRef

[110] Ma J,

[111] Yuan Y

[112] 29.↵
Lan Y,
Han B,
Zhai T, et al
. Clinical application of machine learning-based pathomics signature of gastric atrophy. Front Oncol 2024;14:1289265 doi:10.3389/fonc.2024.1289265 pmid:38476364
CrossRef PubMed

[113] Lan Y,

[114] Han B,

[115] Zhai T, et al

[116] 30.↵
Zhu H,
Hu M,
Ma Y, et al
. Multi-center evaluation of machine learning-based radiomic model in predicting disease free survival and adjuvant chemotherapy benefit in stage II colorectal cancer patients. Cancer Imaging 2023;23:74 doi:10.1186/s40644-023-00588-1 pmid:37537659
CrossRef PubMed

[117] Zhu H,

[118] Hu M,

[119] Ma Y, et al

[120] 31.↵
Murugappan M,
Bourisly AK,
Prakash NB, et al
. Automated semantic lung segmentation in chest CT images using deep neural network. Neural Comput Appl 2023;35:15343–64 doi:10.1007/s00521-023-08407-1 pmid:37273912
CrossRef PubMed

[121] Murugappan M,

[122] Bourisly AK,

[123] Prakash NB, et al

[124] 32.↵
Du H,
Ling L,
Yu W, et al
. Convolutional networks for the segmentation of intravascular ultrasound images: evaluation on a multicenter dataset. Comput Methods Programs Biomed 2022;215:106599 doi:10.1016/j.cmpb.2021.106599 pmid:34974233
CrossRef PubMed

[125] Du H,

[126] Ling L,

[127] Yu W, et al

[128] 33.↵
Lin YC,
Lin G,
Pandey S, et al
. Fully automated segmentation and radiomics feature extraction of hypopharyngeal cancer on MRI using deep learning. Eur Radiol 2023;33:6548–56 doi:10.1007/s00330-023-09827-2 pmid:37338554
CrossRef PubMed

[129] Lin YC,

[130] Lin G,

[131] Pandey S, et al

[132] 34.↵
de Rooij NK,
Greving JP,
Rinkel GJ, et al
. Early prediction of delayed cerebral ischemia after subarachnoid hemorrhage: development and validation of a practical risk chart. Stroke 2013;44:1288–94 doi:10.1161/STROKEAHA.113.001125 pmid:23512975
Abstract/FREE Full Text

[133] de Rooij NK,

[134] Greving JP,

[135] Rinkel GJ, et al

[136] 35.↵
Friedman JA,
Goerss SJ,
Meyer FB, et al
. Volumetric quantification of Fisher Grade 3 aneurysmal subarachnoid hemorrhage: a novel method to predict symptomatic vasospasm on admission computerized tomography scans. J Neurosurg 2002;97:401–07 doi:10.3171/jns.2002.97.2.0401 pmid:12186469
CrossRef PubMed

[137] Friedman JA,

[138] Goerss SJ,

[139] Meyer FB, et al

[140] 36.↵
Huang EP,
O’Connor JP,
McShane LM, et al
. Criteria for the translation of radiomics into clinically useful tests. Nat Rev Clin Oncol 2023;20:69–82 doi:10.1038/s41571-022-00707-0 pmid:36443594
CrossRef PubMed

[141] Huang EP,

[142] O’Connor JP,

[143] McShane LM, et al

[144] 37.↵
Roederer A,
Holmes JH,
Smith MJ, et al
. Prediction of significant vasospasm in aneurysmal subarachnoid hemorrhage using automated data. Neurocrit Care 2014;21:444–50 doi:10.1007/s12028-014-9976-9 pmid:24715326
CrossRef PubMed

[145] Roederer A,

[146] Holmes JH,

[147] Smith MJ, et al

[148] 38.↵
Jia C,
Yang Y,
Xia Y, et al
. Scaling up visual and vision-language representation learning with noisy text supervision. In: Proceedings of the International Conference on Machine Learnin. July 18–24, 2021; Virtual

[149] Jia C,

[150] Yang Y,

[151] Xia Y, et al

[152] 39.↵
Acampa M,
Bongiorno M,
Lazzerini PE, et al
. Increased arterial stiffness is a predictor of delayed ischaemic stroke after subarachnoid haemorrhage. Heart Lung Circ 2021;30:525–30 doi:10.1016/j.hlc.2020.07.016 pmid:32896483
CrossRef PubMed

[153] Acampa M,

[154] Bongiorno M,

[155] Lazzerini PE, et al

[156] 40.↵
Doonan RJ,
Hausvater A,
Scallan C, et al
. The effect of smoking on arterial stiffness. Hypertens Res 2010;33:398–410 doi:10.1038/hr.2010.25 pmid:20379189
CrossRef PubMed

[157] Doonan RJ,

[158] Hausvater A,

[159] Scallan C, et al

[160] 41.↵
Dhar R,
Diringer MN
. The burden of the systemic inflammatory response predicts vasospasm and outcome after subarachnoid hemorrhage. Neurocrit Care 2008;8:404–12 doi:10.1007/s12028-008-9054-2 pmid:18196475
CrossRef PubMed

[161] Dhar R,

[162] Diringer MN

[163] 42.↵
Abdel-Tawab M,
Hasan AA,
Ahmed MA, et al
. Prognostic factors of delayed cerebral ischemia after subarachnoid hemorrhage including CT perfusion: a prospective cohort study. Egypt J Radiol Nucl Med 2020;51:61 doi:10.1186/s43055-020-00180-8
CrossRef

[164] Abdel-Tawab M,

[165] Hasan AA,

[166] Ahmed MA, et al

[167] 43.↵
Woo PY,
Tse TP,
Chan RS, et al
. Computed tomography interobserver agreement in the assessment of aneurysmal subarachnoid hemorrhage and predictors for clinical outcome. J Neurointerv Surg 2017;9:1118–24 doi:10.1136/neurintsurg-2016-012576 pmid:29030464
Abstract/FREE Full Text

[168] Woo PY,

[169] Tse TP,

[170] Chan RS, et al

[171] 44.↵
van der Steen WE,
Marquering HA,
Boers AM, et al
. Predicting delayed cerebral ischemia with quantified aneurysmal subarachnoid blood volume. World Neurosurg 2019;130:e613–19 doi:10.1016/j.wneu.2019.06.170 pmid:31260850
CrossRef PubMed

[172] van der Steen WE,

[173] Marquering HA,

[174] Boers AM, et al

[175] 45.↵
Zijlstra IA,
Gathier CS,
Boers AM, et al
. Association of automatically quantified total blood volume after aneurysmal subarachnoid hemorrhage with delayed cerebral ischemia. AJNR Am J Neuroradiol 2016;37:1588–93 doi:10.3174/ajnr.A4771 pmid:27102313
Abstract/FREE Full Text

[176] Zijlstra IA,

[177] Gathier CS,

[178] Boers AM, et al

Main menu

User menu

Search