Abstract
Objectives
To assess the predictive efficacy of MRI-based radiomics model on the progressive deterioration of neurological function in patients with subacute ischemic perforating artery cerebral infarction through extracting radiomic features within the infarction area on the images of echo planar imaging diffusion weighted imaging (SEPI-DWI), apparent diffusion coefficient (ADC), and T2-weighted image (T2WI) sequences.
Methods
This retrospective study included 188 patients with subacute ischemic perforating artery cerebral infarction. The clinical and head MRI imaging data were collected. We randomly divided patients into training and validation cohort in ratio of 7:3. We used independent sample t test, recursive feature elimination combined with least absolute shrinkage and selection operator (LASSO) regression methods to screen features, four prediction models were developed, and their predictive efficiency was evaluated through ROC curve and decision curve analysis (DCA).
Results
In combined model, the specificity, sensitivity, accuracy, AUC in training cohort were 0.753, 0.761, 0.756, 0.844, respectively, in validation cohort were 0.730, 0.650, 0.702, 0.824, all the values were higher than those of T2WI-model, SEPI-DWI-model and ADC-model, and DCA curve indicated better predictive efficacy in the specific threshold probability range of the combined model.
Conclusion
The MRI-based radiomics model has good predictive efficacy on predicting the progressive neurological dysfunction of patients, and has a certain guiding role in the formulation of personalized treatment for patients with subacute ischemic perforating artery cerebral infarction.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12880-026-02199-1.
Keywords: Progressive cerebral infarction, Magnetic resonance imaging, Radiomics
Introduction
In 1989, Caplan first proposed the concept of “branch atheromatous disease” (BAD) to describe a situation where the origin of a deep cerebral perforating artery is blocked or narrowed due to atherosclerosis [1]. The perforating artery cerebral infarction caused by BAD is a common subtype of ischemic infarction and the infarct sites were mainly in para-body level of lateral ventricles, basal ganglia, pontine and other areas [2, 3]. In diffusion-weighted imaging (DWI) images, the BAD infarction in the lenticulostriate arteries is defined as the lesion diameter exceeding 15 mm or visible in three or more consecutive axial sections; BAD infarction in the paramedian pontine artery is defined as unilateral infarction extending to the base surface of the pons [2, 4]. In recent years, clinicians have found that perforator artery cerebral infarction is non-disabling and mild of clinical symptoms in early stage, but in the course of treatment, such mild clinical symptoms are likely to relapse or worsen, and most of them will be left with dementia, aphasia or paralysis [5, 6]. So, whether patients with perforator artery cerebral infarction suffer from progressive deterioration of neurological function (clinical symptom progression) has been paid more and more attention by clinicians. Therefore, early identification of the progression of clinical symptoms in patients will help to individualize the treatment plan and improve the prognosis of patients.
Radiomics reveals subtle structural changes that cannot be detected by the human eye through the analysis among the region of interest (ROI). The basic principle is that the delineated ROI can give comprehensive quantitative information about the intensity distribution and differences of the pixels within it, which can reveal the basic pathophysiological changes in the tumor or diseased tissue [7–9], with the advantages of objectivity and reproducibility, it has been applied in clinical studies, such as to distinguish tumors with similar appearance, predict the prognosis of disease or treatment response [10–12]. To date, few studies have evaluated the progression of clinical symptoms in patients with perforator artery infarction based on MRI radiomics model. Using this technique can quantify tissue status, develop potential tools to distinguish patients with progressive neurological dysfunction, and assist in formulating therapeutic regimen in these patients with cerebral infarction.
An older definition describes acute stroke as the first 24 h after symptom onset, subacute stroke between day 1 and day 5, and chronic stroke more than 5 days [13]. Another definition describes an early hyperacute phase (0–6 h), a late hyperacute phase (6–24 h), an acute phase (1–7 days), as subacute phase (1–3 weeks) and a chronic phase (>3 weeks) [14]. Khadhraoui E et al. [15] pointed out that the “acute stroke” is that a Diffusion-Weighted Imaging (DWI)-Fluid-Attenuated Inversion Recovery (FLAIR) mismatch and interventions make sense, the “subacute stroke” refers to the stage where secondary complications may occur. If time of symptom onset is less than 4.5 h, patients can proceed immediately to intravenous (IV) thrombolysis. Besides, combining mechanical thrombectomy (MT) for intracranial artery occlusion with IV thrombolysis can extend the treatment window to 24 hours [16]. Therefore, our study intends to establish a multi-parameter MRI radiomics model to evaluate the possibility of progressive deterioration of neurological function in patients with subacute ischemic perforating artery cerebral infarction.
Materials and methods
Patients
We retrospectively collected patients with subacute ischemic perforating artery cerebral infarction from the center 1 from October 2015 to July 2023 and the center 2 from February 2018 to April 2022. The ethics committees approved the study, since this is a retrospective study, the patient’s informed consents were waived.
Inclusion criteria include: (1) Time from onset of stroke to MRI scan no more than 7 days. (2) MRI demonstrated a clear infarction lesion accompanied by neurological dysfunction [15]; (3) Single-lesion ischemic perforating artery cerebral infarction confirmed on DWI based on the diagnostic criteria from existing literature [2, 4]; (4) Imaging and clinical data were complete.
Exclusion criteria include: (1) Transient ischemic attack (TIA) or hemorrhagic cerebral infarction; (2) Head MRI images are incomplete or contain artifacts; (3) Patients who cannot perform NIHSS score because of mental status or osteoarthropathy affecting limb movement. The selecting process of patients with perforator artery cerebral infarction is shown in Fig. 1.
Fig. 1.
The inclusion and exclusion process for patients in this study
Definition of progressive cerebral infarction
Clinically, the degree of progressive deterioration of neurological function in patients with ischemic cerebral infarction was evaluated through National Institutes of Health Stroke Scale (NIHSS) [17, 18]. Progressive cerebral infarction (PCI) is generally considered to be an increase of 2 points or more in NIHSS score during hospitalization compared to score at admission [19]. So, the patients were divided into nonprogressive cerebral infarction (NPCI) and progressive cerebral infarction (PCI). It should be noted that the time of clinical symptom progression in patients should be after the time of the first head MRI examination; otherwise, the patients were not included in this study.
Collection of clinical data
Clinical data mainly include: age, gender, drinking, smoking, NIHSS score at admission and during hospitalization, the time from occurrence of cerebral infarction to first head MRI examination, diastolic and systolic pressure at admission, and the data of TC, TG, HDL, LDL, blood glucose and homocysteine at admission.
The treatment methods included dual antiplatelet therapy or single antiplatelet therapy. During the treatment period, the patients did not receive intravenous thrombolysis or mechanical thrombectomy.
Radiomic features extraction
After hospitalization, all head MR examinations with 1.5T or 3.0T were performed according to standardized protocols. Then, the T2-Weighted Image (T2WI), Single-shot Echo Planar Imaging Diffusion Weighted Imaging (SEPI-DWI) with b-value of 1000s/mm2 and the SEPI-DWI derived Apparent Diffusion Coefficient (ADC) map were used for radiomics feature extraction. The MRI protocols of 1.5T at the center 1 are as follows: (1) T1-Weighted Image: repetition time (TR)/echo time (TE) = 550 ms/8.4 ms, slice thickness/slice spacing = 5.0 mm/1.5 mm, FOV = 240 mm*240 mm; (2) T2WI: TR/TE = 4000 ms/99 ms, slice thickness/slice spacing = 5.0 mm/1.5 mm, FOV = 240 mm*240 mm; (3) SEPI-DWI: TR/TE = 3600 ms/102 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 240 mm*240 mm, b = 0 and 1000s/mm2. The MRI protocols of 3.0T at the center 1 are as follows: (1) T1WI:TR/TE = 1800 ms/9.4 ms, slice thickness/slice spacing = 5.0 mm/0.7 mm, FOV = 256 mm*256 mm; (2) T2WI:TR/TE = 6000 ms/95 ms, slice thickness/slice spacing = 5.0 mm/0.6 mm, FOV = 256 mm*256 mm; (3) SEPI-DWI:TR/TE = 6300 ms/94 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 256 mm*256 mm, b = 0 and 1000 s/mm2. The MRI protocols of 1.5T at the center 2 are as follows: (1) T1WI: TR/TE = 487 ms/14 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 230 mm*230 mm; (2) T2WI: TR/TE = 3690 ms/94 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 230 mm*230 mm; (3) SEPI-DWI:TR/TE = 2900 ms/89 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 230 mm*230 mm, b = 0 and 1000 s/mm2. The MRI protocols of 3.0T at the center 2 are as follows: (1) T1WI: TR/TE = 1500 ms/8.5 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 230 mm*230 mm; (2) T2WI: TR/TE = 3700 ms/96 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 230 mm*230 mm; (3) SEPI-DWI: TR/TE = 5100 ms/100 ms, slice thickness/slice spacing = 5.0 mm/1.0 mm, FOV = 230 mm*230 mm, b = 0 and 1000 s/mm2.
Firstly, Largest layer of lesion was sketched along the boundary manually by one radiologist with more than 3 years working experience through ITK-SNAP (version 3.80). When sketching, the influence of cerebrospinal fluid signals was manually avoided, at the same time, the radiologist was unaware of the clinical characteristics of each patient, then, another radiologist with no less than 10 years’ experience in radiological diagnosis reviewed them, if necessary, the senior radiologist manually modifies the lesion boundary but does not perform segmentation independently. Finally, sketching images and corresponding original images were input together into Dr. Wise Multimodel Research Platform (https://research.deepwise.com) (Deepwise & League of PHD Technology Co., Ltd, Beijing, China, Version 1.6.3.6) which has been reported by previous studies [20, 21] for feature extraction: Firstly, pre-processing was performed, all MRI images were resampled using B-spline interpolation resampling technique so that they were 1 × 1 × 1 mm3 voxels. Secondly, the imaging features of cerebral infarction were obtained including: shape feature; first-order features; Gray-level dependence matrix (GLDM) features; Gray-level co-occurrence matrix (GLCM) features; Gray-level size zone matrix (GLSZM); Gray-level run-length matrix (GLRLM). All features were transformed through square roots, squares, log, logarithms, gradients, wavelets, and local binary pattern (LBP).
Feature selection and model building
We randomly divided patients into training and validation cohort in ratio of 7:3. Before feature selection, the Z-score[(x-μ)/σ] normalization was used to preprocess radiomic features, in this formula, x is feature value, μ is average of feature values of all patients, σ is corresponding standard deviation. feature selection process was as follows: Firstly, the radiomic features with p < 0.05 between two groups were retained through independent sample T-test; Secondly, for the retained features, the least important features were removed through recursive feature elimination (RFE); Lastly, optimal features were selected through least absolute shrinkage and selection operator (LASSO) regression with five-fold cross-validation, also, corresponding coefficients of optimal features were computed; Then, radiomic score (rad-score) was calculated through summing up selected optimal features using weighted coefficients. Finally, the combined prediction model based on support vector machine was established according to the rad-score values of T2WI, DWI and ADC. The study workflow is shown in Fig. 2.
Fig. 2.
Workflow of radiomics model construction
Statistical analysis
The SPSSPRO (https://www.spsspro.com/) and Python (version 3.8.3) were used to analyze data. when data approximate conforming or conforming to normal distribution was represented by mean ± standard (x ± s), while data conforming to skewness distribution was represented by median [interquartile ranges (IQRs)]. Fisher’s exact probability method or Chi-square test for qualitative data, while rank sum test for skewness distribution, t test for normal distribution. Finally, specificity, sensitivity, 95% confidence interval (95%CI), AUC were used for evaluating the predictive efficacy of a specificity model, at the same time, prediction accuracy was evaluated through calibration curve, clinical utility was evaluated through the decision curve analysis (DCA) under different threshold probabilities, p < 0.05 indicates statistical significance.
Results
Population characteristics
There were188 patients with perforator artery cerebral infarction included into study. Table 1 listed clinical characteristics in PCI group and NPCI group. Among this, three clinical indicators of NIHSS on admission, CRP on admission and systolic pressure on admission had statistical significance, which can be included in the multiple regression analysis.
Table 1.
Clinical characteristics of all patients
| Characteristics | NPCI (n = 122) | PCI (n = 66) | P |
|---|---|---|---|
| Sex | 0.527 | ||
| Male | 76 | 38 | |
| Female | 46 | 28 | |
| Age | 63.08 ± 12.47 | 64.68 ± 10.86 | 0.377 |
| NIHSS on admission | 2.20 ± 1.48 | 2.89 ± 1.68 | 0.004 |
| Onset-to-MRI time (day) | 2(1,4) | 2(2,3) | 0.763 |
| CRP on admission (mg/L) | 3.84 ± 6.07 | 6.78 ± 11.02 | 0.019 |
| FIB on admission (g/L) | 2.82 ± 1.04 | 2.95 ± 1.10 | 0.399 |
| D-Dimer on admission (mg/L) | 0.40 ± 0.41 | 0.55 ± 0.75 | 0.082 |
| Systolic pressure on admission (mmHg) | 155.03 ± 20.71 | 163.56 ± 24.98 | 0.013 |
| Diastolic pressure on admission (mmHg) | 88.37 ± 17.84 | 89.38 ± 15.56 | 0.699 |
| Blood glucose on admission (mmol/L) | 6.45 ± 2.74 | 6.93 ± 4.08 | 0.340 |
| Smoking | 0.214 | ||
| No | 90 | 54 | |
| Yes | 32 | 12 | |
| Drinking | 0.794 | ||
| No | 96 | 53 | |
| Yes | 26 | 13 | |
| TG (mmol/L) | 1.56 ± 0.75 | 1.77 ± 1.02 | 0.115 |
| TC (mmol/L) | 4.50 ± 0.99 | 4.47 ± 0.92 | 0.834 |
| LDL (mmol/L) | 2.60 ± 0.81 | 2.48 ± 0.80 | 0.355 |
| HDL (mmol/L) | 1.13 ± 0.33 | 1.13 ± 0.30 | 0.877 |
| HCY (umol/L) | 12.75(10.68, 15.85) | 12.60(10.25, 16.28) | 0.456 |
| Infarct volume (mm3) | 630(472, 942) | 887(651, 1278) | ＜0.001 |
| Medication Therapy | 0.668 | ||
| Mono antiplatelets | 65 | 33 | |
| Dual antiplatelets | 57 | 33 | |
| Location of infarction | 0.063 | ||
| The para-body level of lateral ventricles | 47 | 32 | |
| Basal ganglia | 16 | 5 | |
| Thalamus | 19 | 3 | |
| Pontine | 40 | 26 |
Abbreviations: PCI: Progressive cerebral infarction, NPCI: Non-progressive cerebral infarction, NIHSS: National Institutes of Health Stroke Scale, CRP: C-reactive protein, FIB: Fibrinogen, TG: Triglyceride, TC: Total cholesterol, LDL: Low-density lipoprotein, HDL: High-density lipoprotein, HCY: Homocysteine
Table 2 listed clinical characteristics in training and validation cohorts. 131 patients were in training cohort including 85 in NPCI group (mean age, 63.14 ± 12.59) and 46 in the PCI group (mean age, 65.46 ± 11.28). The validation cohort included 57 patients, with 37 in NPCI group (mean age, 62.89 ± 12.36) and 20 in PCI group (mean age, 62.90 ± 9.87).
Table 2.
Clinical characteristics of patients in training cohort and validation cohort
| Characteristics | Training cohort (n = 131) | Validation cohort (n = 57) | ||||
|---|---|---|---|---|---|---|
| NPCI (n = 85) | PCI (n = 46) | P | NPCI (n = 37) | PCI (n = 20) | P | |
| Sex | 0.048 | 0.048 | ||||
| Male | 54 | 21 | 22 | 17 | ||
| Female | 31 | 25 | 15 | 3 | ||
| Age | 63.14 ± 12.59 | 65.46 ± 11.28 | 0.300 | 62.89 ± 12.36 | 62.90 ± 9.87 | 0.998 |
| NIHSS on admission | 2.32 ± 1.64 | 3.00 ± 1.52 | 0.021 | 1.92 ± 1.01 | 2.65 ± 2.01 | 0.072 |
| Onset-to-MRI time (day) | 2 (1,3) | 2(1.75,3) | 0.651 | 2(1,4) | 2.5(2,3) | 0.891 |
| CRP on admission (mg/L) | 4.02 ± 5.72 | 8.27 ± 12.85 | 0.010 | 3.44 ± 6.89 | 3.36 ± 2.55 | 0.960 |
| FIB on admission (g/L) | 2.88 ± 1.19 | 3.09 ± 1.15 | 0.334 | 2.68 ± 0.55 | 2.65 ± 0.94 | 0.889 |
| D-Dimer on admission (mg/L) | 0.42 ± 0.44 | 0.49 ± 0.72 | 0.484 | 0.36 ± 0.32 | 0.68 ± 0.81 | 0.037 |
| Systolic pressure on admission (mmHg) | 155.47 ± 20.34 | 167.28 ± 25.03 | 0.004 | 154.00 ± 21.79 | 155.00 ± 23.21 | 0.872 |
| Diastolic pressure on admission (mmHg) | 89.65 ± 19.39 | 89.63 ± 16.75 | 0.996 | 85.43 ± 13.40 | 88.80 ± 12.79 | 0.362 |
| Blood glucose on admission (mmol/L) | 6.17 ± 2.58 | 6.94 ± 3.98 | 0.185 | 7.10 ± 3.01 | 6.92 ± 4.41 | 0.852 |
| Smoking | 0.065 | 0.642 | ||||
| No | 62 | 40 | 28 | 14 | ||
| Yes | 23 | 6 | 9 | 6 | ||
| Drinking | 0.236 | 0.315 | ||||
| No | 69 | 41 | 27 | 12 | ||
| Yes | 16 | 5 | 10 | 8 | ||
| TG (mmol/L) | 1.55 ± 0.77 | 1.75 ± 0.98 | 0.188 | 1.58 ± 0.70 | 1.79 ± 115 | 0.397 |
| TC (mmol/L) | 4.57 ± 1.02 | 4.45 ± 0.10 | 0.542 | 4.34 ± 0.91 | 4.50 ± 0.73 | 0.505 |
| LDL (mmol/L) | 2.65 ± 0.83 | 2.48 ± 0.89 | 0.277 | 2.47 ± 0.78 | 2.49 ± 0.56 | 0.950 |
| HDL (mmol/L) | 1.12 ± 0.24 | 1.12 ± 0.28 | 0.951 | 1.15 ± 0.48 | 1.17 ± 0.35 | 0.881 |
| HCY (umol/L) | 13.30(11.05,17.15) | 12.32(10.25,16.15) | 0.209 | 12.20(10.60,14.23) | 13.10(10.13,16.43) | 0.412 |
| Medication Therapy | 0.980 | 0.413 | ||||
| Mono antiplatelets | 46 | 25 | 19 | 8 | ||
| Dual antiplatelets | 39 | 21 | 18 | 12 | ||
| Location of infarction | 0.094 | 0.738 | ||||
| The para-body level of lateral ventricles | 31 | 21 | 16 | 11 | ||
| Basal ganglia | 14 | 5 | 2 | 0 | ||
| Thalamus | 15 | 2 | 4 | 1 | ||
| Pontine | 25 | 18 | 15 | 8 | ||
Abbreviations: PCI: Progressive cerebral infarction, NPCI: Non-progressive cerebral infarction, NIHSS: National Institutes of Health Stroke Scale, CRP: C-reactive protein, FIB: Fibrinogen, TG: Triglyceride, TC: Total cholesterol, LDL: Low-density lipoprotein, HDL: High-density lipoprotein, HCY: Homocysteine
In addition, the cerebral infarction locations were in pontine, thalamus, basal ganglia and para-body level of lateral ventricles. statistical result indicated no statistical significance of the locations between groups (p > 0.05, Tables 1 and 2), among the infarctions located around the lateral ventricle, we identified 3 cases involving the long insular artery (LIA), LIA originates from the island segment of the MCA and is a subtype of white matter medullary arteries (WMMA) [22]. Furthermore, in PCI group, there were 25 patients with infarction lesion located in the right hemisphere and 41 patients with them in the left hemisphere. In NPCI group, there were 46 patients with infarction lesion in the right hemisphere and 76 patients with them in the left hemisphere which showed no statistical difference between the two groups (p > 0.05).
Feature selection
In this study, 1320 radiomic features were obtained based on DWI, ADC and T2WI images respectively. After independent sample T-test, 741 features were retained from T2WI, 462 features from DWI, and 512 features from ADC. Then, after recursive feature elimination and LASSO regression, there were 6 optimal features remaining in T2WI, 4 optimal features remaining in SEPI-DWI, and 11 optimal features remaining in ADC (shown in Fig. 3).
Fig. 3.
Feature selection based on LASSO regression analysis (A–C is the coefficient convergence chart based on T2WI, DWI and ADC group respectively; D–F is the weight chart of optimal feature based on T2WI, DWI and ADC groups respectively)
The calculation formulas of rad-score values of T2WI, DWI and ADC were as follows:
T2WI_radscore = 0.3511 + 0.0117 × lbp-2D_glszm_SizeZoneNonUniformity + 0.0125 × lbp-3D-k_glszm_GrayLevelNonUniformity + 0.0259 × lbp-3D-m2_glcm_ClusterTendency − 0.0003 × lbp-3D-m2_gldm_DependenceVariance − 0.0515×squareroot_glcm_MCC + 0.0187 × wavelet-HH_glrlm_GrayLevelNonUniformityNormalized;
SEPI-DWI_radscore = 0.3511 − 0.0203 × gradient_glcm_MCC + 0.0496 × gradient_ngtdm_Busyness + 0.0635 × lbp-3D-k_glszm_SizeZoneNonUniformity − 0.0354 × original_glcm_MCC;
ADC_radscore = 0.3511 + 0.0014×lbp-2D_glszm_ZoneEntropy − 0.0133 × lbp-3D-m1_glcm_MCC + 0.0077 × lbp-3D-m1_glrlm_LongRunLowGrayLevelEmphasis + 0.0025×lbp-3D-m2_glrlm_GrayLevelNonUniformity + 0.0694×lbp-3D-m2_glszm_SizeZoneNonUniformity − 0.0576 × square_glszm_SmallAreaLowGrayLevelEmphasis − 0.0921 × squareroot_glszm_SizeZoneNonUniformityNormalized + 0.0342 × wavelet-HH_glszm_SmallAreaLowGrayLevelEmphasis − 0.0105 × wavelet-HL_glcm_Imc2 − 0.0025 × wavelet-LL_firstorder_RootMeanSquared − 0.0550 × wavelet-LL_glszm_SmallAreaLowGrayLevelEmphasis.
The clinical indicators NIHSS, CRP, and systolic pressure at admission, which showed statistical differences in Table 1, along with the rad-score values of T2WI, DWI, and ADC were included in multivariate logistic regression analysis. Analysis revealed that NIHSS, CRP, and systolic pressure were not statistically significant (shown in Table 3).
Table 3.
The multivariate logistic regression analysis of clinical indicators and DWI, ADC and T2WI_radscore
| Characteristics | B | Standard Error | OR (95%CI) | P |
|---|---|---|---|---|
| NIHSS on admission | 0.031 | 0.129 | 1.031 (0.800–1.329) | 0.812 |
| CRP on admission (mg/L) | 0.034 | 0.026 | 1.035 (0.983–1.090) | 0.192 |
| Systolic pressure on admission (mmHg) | 0.015 | 0.009 | 1.015 (0.998–1.032) | 0.092 |
| DWI_radscore | 5.850 | 2.255 | 347.221 (4.183–28825.009) | 0.009 |
| ADC_radscore | 3.987 | 1.670 | 53.875 (2.042–1421.509) | 0.017 |
| T2WI_radscore | 9.078 | 3.005 | 8757.586 (24.217–3166982.653) | 0.003 |
Abbreviations: CRP: C-reactive protein, OR: Odds ratio
Evaluation of MRI radiomic models
According to the rad-score, we developed four prediction models: DWI-model, ADC-model, T2WI-model and combined all-model (combined DWI+ADC+T2WI). In training cohort, AUC of DWI-model was 0.769, ADC-model was 0.820, and T2WI-model was 0.776, when features of DWI, ADC and T2WI were combined, AUC of combined all-model increased to 0.844, besides, the sensitivity, specificity, accuracy of combined model was improved reaching 76.1%, 75.3%, 75.6% respectively. In validation cohort, AUC of DWI-model was 0.768, ADC-model was 0.791, and T2WI-model was 0.784, when features of DWI, ADC and T2WI were combined, AUC of combined all-model increased to 0.824, besides, the sensitivity, specificity, accuracy of the combined model was improved reaching 65.0%, 73.0%, 70.2% respectively (shown in Table 4 and Fig. 4), in addition, the DCA curve showed that the combined model had better predictive efficacy within a specific threshold probability range (shown in Fig. 4), the calibration curve demonstrated good consistency between predicted and measured values of combined model (shown in Fig. 5).
Table 4.
Support vector machine prediction model of T2WI, DWI and ADC
| Models | AUC (95%CI) | Accuracy (%) | Sensitivity (%) | Specificity (%) | |
|---|---|---|---|---|---|
| Training cohort | |||||
| (n = 131) | DWI-model | 0.769 (0.680–0.885) | 0.710 | 0.696 | 0.718 |
| ADC-model | 0.820 (0.743–0.883) | 0.718 | 0.717 | 0.718 | |
| T2WI-model | 0.776 (0.684–0.854) | 0.710 | 0.696 | 0.718 | |
| Combined all-model | 0.844 (0.775–0.903) | 0.756 | 0.761 | 0.753 | |
| Validation cohort | |||||
| (n = 57) | DWI-model | 0.768 (0.649–0.874) | 0.667 | 0.450 | 0.784 |
| ADC-model | 0.791 (0.676–0.895) | 0.667 | 0.550 | 0.730 | |
| T2WI-model | 0.784 (0.642–0.899) | 0.667 | 0.600 | 0.703 | |
| Combined all-model | 0.824 (0.720–0.913) | 0.702 | 0.650 | 0.730 | |
Abbreviations: AUC: Area under the curve
Fig. 4.
Radiomic prediction performance of combined all-model (Figure A and B were ROC curves of training cohort and validation cohort to distinguish clinical symptom progression of perforator artery infarction, respectively; Figure C and D were DCA of training and validation cohort, respectively)
Fig. 5.
The calibration curve of the training and validation cohort of combined all-model
Discussion
In this study, we extracted radiomic features from cerebral infarction area based on T2WI, SEPI-DWI, and ADC images and developed models to predict the progression of clinical symptoms in patients with subacute ischemic perforating artery cerebral infarction. We found that combined model based on rad-score value of T2WI+DWI+ADC had a higher AUC in predicting the progressive neurological dysfunction of patients, which could help clinicians make personalized clinical treatment decisions.
Although some studies have assessed the value of radiomics in the diagnosis of ischemic stroke [23], such as prognosis assessment based on ischemic penumbra [24, 25], and prediction of functional outcome [26–28], relatively few have investigated the progression of clinical symptoms in patients with cerebral infarction patients. Guo et al. [29] obtained whole brain dynamic radiomic features (DRF) of patients with cerebral infarction from the 3D images in DSC-PWI time series, and the correlation between DRF and the degree of neurological damage in patients was explored. Their results showed that DRF provided an AUC of 0.853 in evaluating NIHSS scores on admission, suggesting that radiomics can offer valuable insights into the progression of clinical symptoms in patients with perforator artery infarction. Similarly, Wang et al. [30] extracted the radiomic features from cerebral infarction in the pontine region based on DWI and ADC images to establish clinical- radiomics prediction model which had good predictive value in evaluating the early neurological deterioration of isolated acute pontine infarction patients, and the AUC value in training and validation set was 0.966 and 0.920, respectively. The ADC value, as a quantitative parameter of DWI, is widely used in research. Drier et al. [31] indicated that the ADC value can be used to assess the infarct growth and size in the subacute phase. Zhang Y et al. [32] found that magnetic resonance texture analysis on ADC maps was useful to evaluate the clinical function outcomes in patients with unilateral anterior circulation ischemic stroke.
However, Wang J et al. found that age, initial systolic blood pressure, initial NIHSS score and triglycerides showed statistical differences between the two groups, based on these clinical characteristics, a clinical-imaging model was established and presented in the form of nomogram, which was different from our results, we didn’t find the statistical differences of clinical features between the groups which resulted in clinical features not being included to further improve performance of prediction model in our study. We thought this difference may be due to differences in the study population, so the sample size needs to be expanded for further research in subsequent studies.
Traditionally, the assessment of clinical symptom progression in patients with cerebral infarction was mainly based on conventional imaging features. Jiang et al. [33] assessed the presence and total burden of cerebral small vascular disease (CSVD) in patients with single, small subcortical infarction and found that total burden of CSVD was independently associated with clinical symptom progression of single, small subcortical infarction. Our previous study [34] and Oge et al.’s study [35] found that average ADC value in cerebral infarction lesions was closely related to the progressive exacerbation of early neurological deficits in patients with isolated perforator artery cerebral infarction. During the early stage of ischemic cerebral infarction, neurons and glial cells experience energy depletion, resulting in dysfunction of the Na+/K+ pump and water influx into the cells, leading to cytotoxic edema [36] which is manifested as a low signal on the ADC image, based on this, the ADC value can reflect the severity of cytotoxic edema, and indirectly indicate the risk of cell death in the cerebral infarction area [37]. Cytotoxic edema is the principal mechanism that mediates neuronal death [38]. However, identification of these imaging signs is highly subjective. For example, when judging the severity of white matter hypersignal and grading the expanded perivascular space, it is necessary to rely on the diagnostic knowledge and experience stored by radiologists, and the subjective factors are heavy. Compared with the above studies, we propose a feasible method which can provide quantitative feature data objectively and automatically.
Therefore, our finding suggests that the radiomics prediction model has good predictive value in evaluating the progression of clinical symptoms in patients with perforator artery infarction, and has a relatively objective advantage. By establishing a better prediction model, to a certain extent, it can help clinicians to make individualized treatment plan and improve the prognosis of patients with perforator artery infarction.
Notably, the DWI acquisition technique used in this study is SEPI-DWI, which is a commonly employed sequence in brain imaging and can sensitively detect acute and subacute strokes. However, it has limitations in terms of signal-to-noise (SNR) ratio and reduction of artifacts. Compared to standard SEPI-DWI, other DWI techniques also have certain advantages in detecting strokes. Müller SJ et al.’s study [39] found that Diffusion-weighted imaging in stimulated echo acquisition mode (STEAM-DWI) shows significant improvements in CNR, SNR and contrast resolution, and has good sensitivity and specificity for (acute and) subacute infarction. Mavroidis P et al.’s research [40] indicates that Periodically Rotated Overlapping Parallel Lines with Enhanced Reconstruction Diffusion-Weighted Imaging (PROPELLER DWI) not only reduces susceptibility and flow pulsation artifacts, but also achieves higher image quality and lesion clarity, as well as the ability to detect ischemic stroke earlier. Therefore, different DWI acquisition techniques may exhibit certain differences in the display of cerebral infarction lesions, lesion delineation, and extraction of radiomics features, which is a topic worthy of in-depth exploration and requires more clinical research in the future to uncover it.
This study has several limitations. Firstly, the sample size of the patients was small, and they were all individuals with single perforating artery infarction which limited the generalizability of our research results, larger-scale studies and external validations are necessary. Secondly, the clinical indicators showed no significant statistical differences between the groups. In future studies, more clinical variables need to be included to identify meaningful indicators for improving the prediction performance. Thirdly, the location of perforating artery cerebral infarction was not uniform, and to some extent, this may have an impact on the clinical progression outcome, further detailed research is needed. Furthermore, partial volume effect is one of the common technical biases in voxel-based image analysis, especially in very small cerebral infarction lesions which may lead to distortion in the extraction of radiomics features [41], we employed the B-spline interpolation resampling technique for resampling, and preprocessed all MRI images to obtain isotropic radiomics-specific images and methods such as Z-score[(x-μ)/σ] normalization and RFE to conduct radiomics features selection, in order to mitigate the influence of partial volume effect on this study. Ospel JM et al. found that increasing the spatial resolution of DWI would significantly reduce the average measurement error caused by partial volume effect. Finally, brain edema after ischemic stroke may increase the resistance of collateral small arteries or downstream perforating arteries [42] which may lead to deterioration of the patient’s neurological function, besides, the occurrence of hemorrhagic transformation after infarction can also affect progression of neurological function, since this study is a retrospective study, there are limitations in the follow-up of post-infarction edema and hemorrhagic transformation, further exploration is needed in subsequent retrospective and prospective studies.
Conclusions
In summary, we developed a radiomics-based model using T2WI, DWI, and ADC images to predict progression of clinical symptoms in patients with subacute ischemic perforating artery cerebral infarction. The combined model demonstrated higher AUC and predictive performance compared to previous approaches. Our findings support the utility of radiomics in providing objective, quantitative data for predicting clinical outcomes, offering advantages over traditional imaging assessments.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Acknowledgements
We thank all scientists who have contributed to science.
Author contributions
The authors would like to thank WJY for writing – editing & original draft, Methodology. WWS for writing – review, Methodology, Conceptualization. JFL for data curation. HQ for conceptualization. ZXZ for project administration. LPZ for investigation. ZJH for project administration. ZJC for writing – review & editing, Funding acquisition. MSX for writing – review & editing, Funding acquisition, Supervision.
Funding
This work was supported by the Zhejiang Chinese Medical University [grant number 2022FSYYZZ01], The Medical Health Science and Technology Project of Zhejiang Provincial Health Commission [grant number 2024KY1210] for the execution and analysis.
Data availability
The datasets used and analyzed in the current study are available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
The study was approved by the Institutional Review Board of The First Affiliated Hospital of Zhejiang Chinese Medical University [No. 2022-KLS-197-01], and the requirement of individual consent for this retrospective analysis was waived. All procedures performed in studies were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Caplan LR. Intracranial branch atheromatous disease: a neglected, understudied, and underused concept. Neurology. 1989;39(9):1246–50. [DOI] [PubMed] [Google Scholar]
- 2.Sun S, Wang Y, Wang Y, et al. Lipid and hyperglycemia factors in first-ever penetrating artery infarction, a comparison between different subtypes. Brain Behav. 2017;7(6):e00694. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Gore M, Bansal K, Khan Suheb MZ, et al. Lacunar Stroke. In: StatPearls. Treasure Island (FL): StatPearls Publishing; 2023 Jan 16. [PubMed]
- 4.Kanazawa K, Miyamoto N, Hira K, et al. Relative score of early neurological deterioration in perforator artery infarction: a retrospective study. BMC Neurol. 2024;24(1):298. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Li S, Ni J, Fan X, et al. Study protocol of branch atheromatous disease-related stroke (BAD-study): a multicenter prospective cohort study. BMC Neurol. 2022;22(1):458. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Ginex V, Gilardone G, Viganò M, et al. Interaction between recovery of motor and language abilities after stroke. Archives Phys Med Rehabil. 2020;101(8):1367–76. [DOI] [PubMed] [Google Scholar]
- 7.Sarioglu O, Sarioglu FC, Capar AE, et al. The role of CT texture analysis in predicting the clinical outcomes of acute ischemic stroke patients undergoing mechanical thrombectomy. Eur Radiol. 2021;31(8):6105–15. [DOI] [PubMed] [Google Scholar]
- 8.Rizzo S, Botta F, Raimondi S, et al. Radiomics: the facts and the challenges of image analysis. Eur Radiol Exp. 2018;2(1):36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–77. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Zhang S, Chiang GC, Magge RS, et al. Texture analysis on conventional MRI images accurately predicts early malignant transformation of low-grade gliomas. Eur Radiol. 2019;29(6):2751–59. [DOI] [PubMed] [Google Scholar]
- 11.Miao L, Cao Y, Zuo L, et al. Predicting pathological complete response of neoadjuvant radiotherapy and targeted therapy for soft tissue sarcoma by whole-tumor texture analysis of multisequence MRI imaging. Eur Radiol. 2022. [DOI] [PMC free article] [PubMed]
- 12.Li J, Fu S, Gong Z, et al. MRI-based texture analysis of infrapatellar fat pad to predict knee osteoarthritis incidence. Radiology. 2022;304(3):611–21. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Birenbaum D, Bancroft LW, Felsberg GJ. Imaging in acute stroke. West J Emerg Med. 2011;12(1):67–76. [PMC free article] [PubMed] [Google Scholar]
- 14.Allen LM, Hasso AN, Handwerker J, et al. Sequence-specific MR imaging findings that are useful in dating ischemic stroke. Radiographics. 2012;32(5):1285–97. [DOI] [PubMed] [Google Scholar]
- 15.Khadhraoui E, Müller SJ. Lack of standardization in the nomenclature of dating strokes or the desperate search for a common language. Int J Emerg Med. 2025;18(1):14. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Ho JP, Powers WJ. Contemporary management of acute ischemic stroke. Annu Rev Med. 2025;76(1):417–29. [DOI] [PubMed] [Google Scholar]
- 17.Xu XQ, Shen GC, Ma G, et al. Prognostic value of post-treatment fluid-attenuated inversion recovery vascular hyperintensity in ischemic stroke after endovascular thrombectomy. Eur Radiol. 2022;32(12):8067–76. [DOI] [PubMed] [Google Scholar]
- 18.Chalos V, van der Ende NAM, Lingsma HF, et al. National Institutes of health stroke scale: an alternative primary outcome measure for trials of acute treatment for ischemic stroke. Stroke. 2020;51(1):282–90. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Poh KW, Er CK, Hoh WH, et al. Neurological deterioration and its risk score in total anterior circulation infarct. Clin Neurol Neurosurg. 2020;19(1):105684. [DOI] [PubMed] [Google Scholar]
- 20.Jiang Y, Li W, Huang C, et al. A computed tomography-based radiomics nomogram to preoperatively predict tumor necrosis in patients with clear cell renal cell arcinoma. Front Oncol. 2020;10:592. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Ding H, Li J, Jiang K, et al. Assessing the inflammatory severity of the terminal ileum in Crohn disease using radiomics based on MRI. BMC Med Imag. 2022;22(1):118. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Tamura A, Kasai T, Akazawa K, et al. Long insular artery infarction: characteristics of a previously unrecognized entity. AJNR Am J Neuroradiol. 2014;35(3):466–71. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Ortiz-Ramón R, Valdés Hernández MDC, González-Castro V, et al. Identification of the presence of ischaemic stroke lesions by means of texture analysis on brain magnetic resonance images. Comput Med Imaging Graph. 2019;74:12–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Tang TY, Jiao Y, Cui Y, et al. Penumbra-based radiomics signature as prognostic biomarkers for thrombolysis of acute ischemic stroke patients: a multicenter cohort study. J Neurol. 2020;267(5):1454–63. [DOI] [PubMed] [Google Scholar]
- 25.Guo Y, Yang Y, Cao F, et al. Novel survival features generated by clinical text Information and radiomics features may improve the prediction of ischemic stroke outcome. Diagnostics (Basel). 2022;12(7):1664. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Wang H, Sun Y, Ge Y, et al. A clinical-radiomics nomogram for functional outcome predictions in ischemic stroke. Neurol Ther. 2021;10(2):819–32. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Zhou Y, Wu D, Yan S, et al. Feasibility of a clinical-radiomics model to predict the outcomes of acute ischemic stroke. Korean J Radiol. 2022;23(8):811–20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Wu TC, Liu YL, Chen JH, et al. Prediction of poor outcome in stroke patients using radiomics analysis of intraparenchymal and intraventricular hemorrhage and clinical factors. Neurol Sci. 2023;44(4):1289–1300. [DOI] [PubMed]
- 29.Guo Y, Yang Y, Cao F, et al. A focus on the role of DSC-PWI dynamic radiomics features in diagnosis and outcome prediction of ischemic stroke. J Clin Med. 2022;11(18):5364. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Wang J, Fu K, Wang Z, et al. MRI-based clinical-radiomics nomogram to predict early neurological deterioration in isolated acute pontine infarction: a two-center study in northeast China. BMC Neurol. 2024;24(1):39. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Drier A, Tourdias T, Attal Y, et al. Prediction of subacute infarct size in acute middle cerebral artery stroke: comparison of perfusion-weighted imaging and apparent diffusion coefficient maps. Radiology. 2012;265(2):511–17. [DOI] [PubMed] [Google Scholar]
- 32.Zhang Y, Zhuang Y, Ge Y, et al. MRI whole-lesion texture analysis on ADC maps for the prognostic assessment of ischemic stroke. BMC Med Imag. 2022;22(1):115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Jiang J, Huang X, Zhang Y, et al. Total MRI burden of cerebral vessel disease correlates with the progression in patients with acute single small subcortical strokes. Brain Behav. 2019;9(1):e01173. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Yu W, Yang J, Liu L, et al. The value of diffusion weighted imaging in predicting the clinical progression of perforator artery cerebral infarction. NeuroImage Clin. 2022;35:103117. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Oge DD, Topcuoglu MA, Arsava EM. Apparent diffusion coefficient signature of ischemic tissue predicts neurological progression in isolated pontine infarcts. Eur Stroke J. 2022;7(1):66–70. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Durukan A, Tatlisumak T. Acute ischemic stroke: overview of major experimental rodent models, pathophysiology, and therapy of focal cerebral ischemia. Pharmacol Biochem Behav. 2007;87(1):179–97. [DOI] [PubMed] [Google Scholar]
- 37.Momosaka D, Togao O, Kikuchi K, et al. Correlations of amide proton transfer-weighted MRI of cerebral infarction with clinico-radiological findings. PLoS ONE. 2020;15(8):e0237358. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Dreier JP, Lemale CL, Kola V, et al. Spreading depolarization is not an epiphenomenon but the principal mechanism of the cytotoxic edema in various gray matter structures of the brain during stroke. Neuropharmacology. 2018;134(Pt B):189–207. [DOI] [PubMed] [Google Scholar]
- 39.Müller SJ, Khadhraoui E, Kube JMV, et al. Diagnostic value of diffusion-weighted STEAM-MRI in ischemic stroke. Eur J Radiol. 2021;139:109677. [DOI] [PubMed] [Google Scholar]
- 40.Mavroidis P, Giankou E, Papaioannou M, et al. Comparison between EPI DWI and PROPELLER DWI in brain MR imaging. Curr Probl Diagn Radiol. 2024;53(1):73–80. [DOI] [PubMed] [Google Scholar]
- 41.Ospel JM, Jaffray A, Schulze-Zachau V, et al. Spatial resolution and the magnitude of infarct volume measurement error in DWI in acute ischemic stroke. AJNR Am J Neuroradiol. 2020;41(5):792–97. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Rocha M, Jovin TG. Fast versus slow progressors of infarct growth in large vessel occlusion stroke: clinical and research implications. Stroke. 2017;48(9):2621–27. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets used and analyzed in the current study are available from the corresponding author on reasonable request.