This research was approved by the ethics committee, and the need of obtaining informed consent of patients was waived owing to the retrospective design of the study.

Magnetic resonance imaging data of 152 patients from Medical Center A (ZPP hospital) and 80 patients from Medical Center B (LSP hospital) were collected in this study. The labeled names of all patients of the Medical Center A dataset were listed in alphabetical order and divided into two sets: the division formed a training set (n = 105) and a testing set (n = 47) in the ratio of 7:3. The database of Medical Center B patients was used as an external validation dataset. Prins et al. (2004) proposed a visual rating scale in 2004; based on this scale, all patients were divided into WMH progression group (n = 57 in A and n = 31 in B) and non-progression group (n = 95 in A and n = 49 in B). Periventricular WMH (PWMH) and deep WMH (DWMH) were compared independently. PWMH was defined as WMH within 10 mm from the ventricle surface. WMH away from the ventricle surface 10 mm is defined as DWMH. Scores of −3 to +3 were given according to the progression or decrease in PWMH in the former horn, body, and posterior horn, as shown in Figure 1. Scores of −4 to +4 were given according to the progression or decrease in DWMH in different brain regions. WMH progression was defined when the total score was ≥1. WMHs were graded, and their volume was quantified using FLAIR images. Clinical information on various aspects, such as vascular risk factors, clinical course, age, and sex, was obtained from the medical records of the picture archiving and communications system.

We included patients who (1) had a clinical diagnosis of minor strokes or transient ischemic attacks, (2) underwent more than two MRI examinations on the same machine within an interval of 2–3 years, (3) were older than 60 years at the first examination, and (4) had visible WMH at baseline. We excluded patients who (1) had acute vascular lesions, such as ischemic stroke (except for lacunar infarction) or intracranial hemorrhage; (2) had non-vascular white matter lesions, such as immunologic demyelination, metabolic encephalopathy, poisoning, and infection; (3) had other intracranial lesions, including Alzheimer’s disease, Parkinson’s disease, craniocerebral trauma, or tumor; (4) had incomplete clinical data; (5) had incomplete imaging data; and (6) had imaging data with motion or machine artifacts. Figures 2, 3 show the flowchart summarizing participant recruitment and building radiomic models.

Brain MRI scans in medical centers A and B were performed using a 3.0 T MRI scanner with an eight-channel head coil (Discovery MR 750, GE Healthcare, Chicago, IL, United States). The routine sequences were as follows: axial T₁WI, T₂WI, FLAIR, and DWI. We used axial FLAIR to observe and segment the identified WMH based on the following parameters: TR = 9,000 ms, TE = 120 ms, field of view (FOV) = 220 mm × 220 mm, matrix = 256 × 256, and section thickness = 5 mm, and inter-slice gap = 1.5 mm. We used T₁WI for segmenting the white matter with TR = 1,750 ms, TE = 24 ms, FOV = 220 mm × 220 mm, section thickness = 5 mm, and inter-slice gap = 1.5 mm. T₂WI: TR = 9,823 ms, TE = 101 ms, FOV = 220 mm × 220 mm, section thickness = 5 mm, and inter-slice gap = 1.5 mm. DWI: TR = 3,071 ms, TE = minimum, b = 0, 1,000 s/mm², FOV = 220 mm × 220 mm, inter-slice gap = 1.5 mm, and section thickness = 5 mm.

All MRI scan sequences were converted to NIfTI files (.nii.gz) and preprocessed for normalization. Different imaging sequences were co-registered to the same anatomical template; then, they were interpolated to the same resolution (1 mm × 1 mm × 1 mm) and skull-stripped (Rohlfing et al., 2010; Bakas et al., 2017). The noise in the images was reduced using Gaussian filtering; then, for reducing external interference factors, magnetic field migration correction was performed. Then, by downsampling each image into 32 bins, the image grayscale intensity level was discretized and normalized for noise reduction. Given these fixed bins values and numbers, the grayscale range of the image is divided into equally spaced intervals. Thus, the grayscale values reflected the size and intensity resolution of the discretized bins (i.e., there are four sized bins per grayscale).

All MRI images and ROIs were imported into the AGK software to extract radiomic features. The radiomic features were calculated, including histogram, texture, form factor, gray-level co-occurrence matrix (GLCM), run-length matrix (RLM), and gray-level size zone matrix (GLSZM). The extracted texture features were standardized, which removed the unit limits of the data of each feature and converted it into a dimensionless pure value. This allowed the indexes of different units or orders to be compared and weighted. For details, see Supplementary Materials.

Based on the training set, we performed analysis of variance (ANOVA) of the extracted features. For feature dimensionality reduction, the analysis of ANOVA + Mann–Whitney U-test, correlation analysis, and gradient boosting decision tree were sequentially performed. See Supplementary Materials for details. The five machine learning methods, namely Bayes, the random forest, the logistic regression, the support vector machine classifiers (SVM), and the k-nearest neighbor (KNN), were used to build models, and the best modeling method was selected through comparison. The test set, the training set, and the external verification set were used to verify the predictive efficiency, including calibration efficiency, net value, and diagnostic accuracy, which were estimated using the calibration curve, the decision curve analysis (DCA), and receiver operating characteristic (ROC) curve.

For eliminating the sulcus and gyrus, the ROI WMHp was first manually adjusted by the physician XDH. One month later, the physicians XDH and YS manually eliminated these again on 30 randomly selected subjects. The intra-observer correlation coefficient was calculated based on the first measurement of the physicians XDH and YS. The interobserver correlation coefficient was calculated based on the two measurements acquired by the physician XDH.

We performed our statistical analyses using SPSS 20.0 (IBM, Chicago, IL, United States). Comparisons of clinical and imaging characteristics were performed using a T-test, Mann–Whitney U-test, or chi-square test. Moreover, we performed univariate logistic regression analyses on each potential predictor variable associated with WMH progression, including age, sex, imaging factors, and vascular risk factors. Thereafter, to construct combined prediction models, factors with marginal significance (P < 0.1) in univariate logistic regression were included in multivariable logistic regression. The pairwise correlation among clinical features, the radiomic score (rad-score), and the speed of WMH progression were calculated using the Spearman’s analysis. The values of P ≤ 0.05 were considered to indicate statistical significance.

Table 1 shows the imaging features and demographic characteristics of the participants. In Medical Center A, the median age of patients with WMH progression was significantly higher than for those without WMH progression (74 years vs. 68 years, P = 0.004). In Medical Center B, the average age of patients was higher in the WMH progression group than in the non-progression group; however, the difference was not statistically significant (73.52 years vs. 71.65 years, P = 0.327). In both medical centers, the difference in the volume and speed of WMH progression was statistically significant between the two groups (all P < 0.05). We found no significant difference in imaging features and demographic characteristics between medical centers A and B (all P > 0.05).

For constructing radiomic models after feature dimensionality reduction, the optimal features were selected, including 12 features in the ROI WMHp and 7 features in the ROI WBWM. The area under the curve (AUC) value was higher for the logistic regression model than for other machine learning methods. In line with this finding, the models were built using the logistic regression classifier (Figure 4). The rad-score was calculated using the formula for the features. The rad-score was found to be significantly different between the progression and non-progression groups in two ROIs (all P < 0.05; Table 2). Additional information about the formulas is shown in the Supplementary Material. The predictive efficacies (represented as the AUC) of the ROI WBWM were 0.725, 0.693, and 0.691 for the training group, testing group, and external validation group, respectively. Similarly, the predictive efficacies of the ROI WMHp were 0.791, 0.768, and 0.767 for the training group, testing group, and external validation group, respectively. Figures 5–7 show the diagnostic accuracy, the calibration efficiency, and the net value of models, which were evaluated using the ROC, the Hosmer–Lemeshow test, and DCA, respectively.

After stepwise logistic regression analysis, age, hypertension, and the rad-score of the ROI WBWM were the independent factors of WMH progression (Table 3). We used these factors to construct the combined model in WBWM. The AUCs were 0.779, 0.716, and 0.673 in the training group, testing group, and external validation group, respectively (Table 4). In addition, the combined model in WBWM was constructed using the age and the rad-score of the ROI WMHp (Table 3). The AUCs were 0.793, 0.774, and 0.777 in the training group, testing group, and external validation group, respectively (Table 4).

The speed of WMH progression was positively correlated with the rad-score from the ROI WMHp and age (r = 0.49 and r = 0.15, respectively). In addition, in our sample, we found a weak correlation between hypertension and coronary heart disease, hypertension and diabetes, and hyperlipidemia and smoking (r = 0.14, r = 0.16, and r = 0.18, respectively). Women were more likely to develop diabetes than men (r = −0.16), as shown in Figure 8.

The intra- and interobserver correlation coefficients of segmenting the ROI WMHp were 0.846 and 0.885, respectively.