Radiomics machine learning based on asymmetrically prominent cortical and deep medullary veins combined with clinical features to predict prognosis in acute ischemic stroke: a retrospective study

Participants

The clinical and imaging data of patients who were diagnosed with AIS and hospitalized in the Department of Neurology of Liaoning Provincial People’s Hospital between January 2018 and June 2023 were retrospectively analyzed. The functional outcomes of the patients were evaluated with the modified Rankin scale (mRS). The study adhered to Helsinki principles for studies involving human subjects and received approval from the ethical committee of the People’s Hospital of Liaoning Province (approval no. (2024) H004). Since this was a retrospective study, additional informed consent was waived.

The inclusion criteria were as follows: (I) the diagnostic criteria of the Chinese guidelines for the diagnosis and treatment of acute ischemic stroke 2023 (Chinese Society of Neurology et al., 2024); and (II) head MR imaging data obtained within 72 h after onset, including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), T2 fluid attenuated inversion recovery (FLAIR) imaging, MRA, DWI, and SWI.

The exclusion criteria were as follows: (I) other abnormal brain MRI results, such as cerebral hemorrhage, brain space-occupying lesions, head trauma, and cerebrovascular malformations; (II) infarction combined with hemorrhage; (III) MR images with severe artifacts that would affect the results of the image analysis; (IV) previous stroke with sequelae and a basic mRS score ≥1; and (V) loss to follow-up.

The patient screening and research workflows are shown in Figs. 1 and 2. To balance the data during model training, we ensured a 1:1 ratio between groups. A total of 150 AIS patients (109 males and 41 females; aged 65.13 ±  10.99 years) were included in this study, and their outcomes were grouped according to the mRS score. Patients with a good outcome were defined as those with an mRS score ranging from 0–2 points, and those with a poor outcome were defined as those with an mRS score ≥3 points. Thus, a total of 75 patients were included in the poor outcome group and the good outcome group. In the radiomic study, to verify the performance of the machine learning model, the patient data were randomly allocated to the training and test sets at a 7:3 ratio (n = 105 and n = 45, respectively).

Image data acquisition

All patients were scanned with a Discovery 750 3.0T MRI scanner equipped with an 8-channel coil (GE HealthCare, Chicago, IL, USA). The patient was placed in the supine position with the head facing forward and scanned parallel to the anterior commissure-posterior commissure line from the foramen magnum to the cranial vault. The scan sequences included conventional T1WI, T2WI, FLAIR, DWI and SWI sequences. The SW images were subjected to minimum intensity projection to obtain the mIP map. The scan parameters of each sequence are shown in Table S1. All scan data were uploaded to an ADW 4.7 workstation (GE HealthCare, Chicago, IL, USA) for postprocessing and analysis.

Clinical data collection

Data were collected as previously described in Li et al. (2023). Specifically age at admission, sex, systolic and diastolic blood pressure, fasting glucose, cholesterol, total homocysteine (tHcy), National Institutes of Health Stroke Scale (NIHSS) score at admission, and mRS score three months post discharge. Imaging indicators included the DMV score, DMV symmetry, cerebral cortical vein quantification score, infarct area, and white matter hyperintensity (WMH) score.

(1) DMV score: The DMVs that drain the frontal lobe and anterior parietal lobe white matter are distributed in a vertical fan shape next to the anterior horn of the lateral ventricle. The DMVs that drain the posterior parietal lobe and the occipital lobe white matter are distributed throughout the posterior horn or triagonal area of the lateral ventricle, primarily in the upper and lower planes relative to the lateral paraventricular plane. Therefore, we divided the planes above and below the lateral paraventricular plane into six regions (Fig. 3). The DMV score was calculated as the score for each area, determined as follows (Li et al., 2023; Wang et al., 2024; Zhang et al., 2017): 0 points = each vein was continuous, and the signal was uniform; 1 point = each vein was continuous, but one or more veins had uneven signals; 2 points = one or more veins were discontinuous, manifesting as a punctate low signal; 3 points = none of the veins were continuous.

(2) Determination of DMV symmetry: DMVs were considered symmetrical if they were observed in the same area near the lateral ventricles with similar degrees of dilation and differed between the hemispheres by no more than five veins (Li et al., 2023).

(3) White matter hyperintensity (WMH) evaluation: On FLAIR imaging, WMH evaluation was performed according to the Fazekas scale (0–6 points), as previously described (Wang et al., 2022).

(4) Infarct area: The area was measured and classified on the slice showing the maximum infarct area on DWI: (I) lacunar infarction: ≤1.5 cm²; (II) small infarction: 1.6–3.0 cm²; (III) moderate infarction: 3.1–5.0 cm² and occupying less than one lobe of the brain; (IV) large-area infarction: occupation of more than one lobe of the brain or an infarct area larger than 5.0 cm².

(5) Quantitative cerebral cortical vein score: The prominent cortical veins were scored on the SWI mIP images with reference to the Alberta Stroke Project Early CT Score (ASPECTS). Specifically, the brain is divided into the following areas: L = lenticular nucleus; M1 = anterior middle cerebral artery (MCA) cortex; M2 = MCA cortex on the lateral side of the insula; M3 = cortical area at the posterior part of the MCA; M4–M6: cortical areas above the respective M1–M3 areas. Each area is given a score of 1 point, with a maximum prominent vessel sign (PVS) score of 7 points. The value of normal cortical veins is 0, whereas a score of seven indicates diffuse dilatation of the cortical veins. Figure 4 shows a graph of asymmetrically prominent cortical veins.

Region of interest segmentation

The region of interest (ROI) was manually segmented slice by slice by two junior radiologists with 3 and 5 years of experience in 3D Slicer software (version 5.2.2, https://www.slicer.org/). The segmentation results were reviewed and validated by a senior radiologist with 15 years of experience. On each consecutive imaging slice from the SWI sequence, the ROI was delineated along the contour of each DMV to cover the entire volume of the vein. Up to five DMV ROIs annotated with their morphological characteristics were delineated in each of the six regions according to the DMV score, as shown in Figs. 5A–5E. On each consecutive imaging slice of the SWI mIP images, each ROI was delineated according to the profile of the APCV to cover the entire volume of the vein (Figs. 5F–5J). These ROIs were subsequently used for radiomic feature extraction and screening.

Data preprocessing

The data were randomly allocated at a ratio of 7:3 to a training set and a test set. All the data in the training set were used to train the predictive model, whereas those in the test set were used to independently evaluate the performance of the model. Owing to scanner or acquisition differences, unlike other schemes, spatial normalization is typically used to reduce the effect of differences in voxel spacing. To solve this problem, we used a fixed-resolution resampling method in our experiments, resampling all the images to a size of 1 × 1 ×1 mm, Z score normalization, Laplacian–Gaussian filtering (σ = 2–5) and wavelet transform to extract multiscale features.

Feature extraction

Radiomic features were extracted from each ROI via the PyRadiomics package (version 3.1.0, https://pyradiomics.readthedocs.io) in Python (version 3.7.0). The MR images and corresponding ROIs were simultaneously used to extract radiomic features, which could be divided into three groups: (I) geometric features; (II) intensity features; and (III) texture feature. To address the scale heterogeneity among the artificial radiometric features, the features were normalized via the Z score method. To obtain high-throughput features, nonlinear intensity filtering was performed at the voxel level with the Laplacian−Gaussian filter (with sigma values set to 2, 3, 4, and 5) for first-order statistics and eight wavelet transform algorithms (LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH) for texture features.

Feature selection

First, the Mann–Whitney U test was used to identify radiomic features in the training set that were significantly different between the groups (P < 0.05). Next, Person’s rank correlation analysis was used to calculate the correlation coefficients between the features, and one feature from a pair of features with a correlation coefficient greater than 0.9 was retained. The minimum redundancy–maximum relevance (mRMR) algorithm was used to retain features with high correlations and low redundancy. Finally, the least absolute shrinkage and selection operator (LASSO) algorithm was used for further feature screening. The optimal λ value was determined with 10-fold cross-validation and the minimum standard, and the features with nonzero coefficients were retained.

LASSO regression selects features by minimizing the following loss function: ${\displaystyle \underset{\beta }{min}\left(\frac{1}{2N}\sum _{i=1}^N{\left(y_i-{\beta }^T{X}_i\right)}^2+\lambda \parallel \beta {\parallel }_1\right)}$ where N is the number of samples, λ is the regularization parameter, and the optimal value is determined via 10-fold cross-validation (Fig. 6).

Clinical factor analysis

To select clinical factors significantly associated with patient outcomes, univariable analysis was conducted on 13 clinical factors to identify those with significant differences between the good outcome group and the poor outcome group, after which multivariable regression analysis was performed to identify independent risk factors for poor outcomes in AIS patients.

Model construction

Statistical analysis

All the data were analyzed with the Python statamodels package (version 0.13.2), and a P value < 0.05 was considered to indicate statistical significance. For continuous variables, we used the t test or the Mann–Whitney U test to analyze differences among groups; for categorical variables, we used the chi-square test or Fisher’s test for between-group comparisons. For screening baseline clinical characteristics, univariable analysis was performed first, followed by multivariable regression analysis. For radiomic characteristics, we used the Mann–Whitney U test and Pearson’s rank correlation analysis, the mRMR algorithm, the LASSO algorithm, and the cross-validation method for screening. The screened features were input into the SVM algorithm to construct a risk model. The DeLong test was used to compare performance metrics between the models. Moreover, we generated receiver operating characteristic (ROC) curves to evaluate the prediction model. The area under the curve (AUC), sensitivity (Sen), specificity (Spe), accuracy (ACC), confidence interval (CI) and precision were also calculated and analyzed.

Clinical characteristics

Following the application of the inclusion and exclusion criteria, a total of 150 patients were included in the study, with a mean age of 65.13 ± 10.99 years. There were 109 males and 41 females, including 75 patients in the good outcome group and 75 patients in the poor outcome group. The baseline clinical data of the patients in the two groups and the results of their comparisons are shown in Table 1. Table 2 lists the results of multivariable regression analysis of the clinical characteristics, NIHSS score P = 0.000, DMV score P = 0.004, infarct size P = 0.041, prominent cortical vein score P = 0.747, and DMV symmetry P = 0.394 in the training set.

Radiomics model

Radiomic features were extracted from the DMV- and APCV-delineated ROIs. A total of 1,197 features were extracted, including 234 first-order features, 14 shape features and 949 texture features. A total of 431 features were screened with the t test or the Mann–Whitney U test and Pearson’s correlation test. The use of the mRMR algorithm resulted in the retention of 20 features with high correlation and low redundancy. Finally, through LASSO regression, tenfold cross-validation was performed using the minimum criterion to determine the optimal λ value (Figs. 6A and 6B). Ultimately, nine DMV and APCV radiomic features with nonzero coefficients were selected (Fig. 6C). Finally, the selected features were input into the SVM algorithm, and the radiomic feature model was constructed through fivefold cross-validation. The ROC curves of the model in the training set (AUC =0.986) and the test set (AUC = 0.976) are shown in Figs. 7A–7D. Table 3 shows that in the test set, Sen = 0.952, Spe = 0.955, PPV = 0.952, and NPV = 0.955. In addition to the AUC, sensitivity, and specificity, F1 scores were calculated to assess model balance comprehensively. Fivefold cross-validation revealed a mean AUC of 0.992 (standard deviation ± 0.006) for the combined model.

Clinical model

Multivariate regression analysis of the clinical characteristics that differed between the groups revealed that the NIHSS score, DMV score and infarct area were independent risk factors for poor AIS patient outcomes. Therefore, these three factors were selected as inputs into the SVM algorithm to construct the clinical model. The ROC curves of the model in the training set (AUC = 0.818) and test set (AUC = 0.816) are shown in Figs. 7A and 7D. In the test set, Sen=0.714, Spe=0.864, PPV=0.833, and NPV=0.76, as shown in Table 3.

Combined model

By integrating the radiomics model and the clinical model, a joint model was established. Table 3 shows the performance indicators of the joint model, with AUC = 0.998 in the training set and AUC = 0.996 in the test set. The ROC curves are shown in Figs. 7A and 7D. Moreover, in the test set, Sen = 0.905, Spe = 1.000, PPV = 1.000, and NPV = 0.917. Figure 7G shows the nomogram of the joint model incorporating clinical features and APCV-DMV radiomic features. Here, the scores of the individual elements are added to obtain the total score. The predicted risk values shown correspond to the probabilities of a poor outcome for AIS patients. Figure 7B and e shows the calibration curves for the model in the training and test sets. The plots show that in both sets, the consistency between the predicted and observed outcomes of AIS patients was consistent.

The individual and joint models were also evaluated via decision curve analysis (DCA). As shown in Figs. 7C and 7F, DCA of the clinical model, the APCV-DMV radiomic model and the joint model revealed that the joint model had the greatest net clinical benefits in predicting the outcomes of AIS patients before treatment. Similarly, the DeLong test suggested that the joint model could best predict the outcomes of AIS patients (P < 0.05). The AUC value for predicting the outcomes of AIS patients was significantly greater with the joint model than with the other two models in both the training set and the test set (Table S3).