A total of 69 patients in the chronic phase of subcortical stroke were recruited from Huashan Hospital and underwent neuroimaging at the Shanghai Key Laboratory of Magnetic Resonance. Inclusion criteria were: (1) first-ever subcortical stroke; (2) duration of illness > 3 months; (3) Montreal Cognitive Assessment (MoCA) score > 23, indicating preserved global cognition; and (4) right-handedness. Exclusion criteria were: (1) contraindications to MRI; (2) bilateral stroke lesions; and (3) comorbid neurological or psychiatric disorders unrelated to stroke. Cognitive function was assessed using the MoCA, administered by an experienced therapist at admission. UL motor function was evaluated using the UL-FMA scale. The study was conducted in accordance with the Declaration of Helsinki, approved by the Review Board of Ethics Committee of Huashan Hospital, and registered at the Chinese Clinical Trial Registry (ChiCTR-TRC-08003005). Written informed consent was obtained from all participants.

To investigate resting-state functional connectivity alterations associated with UL motor impairment severity, patients were categorized into two groups based on their UL-FMA scores: those scoring > 20 were classified as having mild-to-moderate impairment (Mild/Moderate Stroke Patients, MSP; n = 29), and those with scores ≤ 20 as having severe impairment (Severe Stroke Patients, SSP; n = 40) (9, 13, 31). Demographic comparisons between groups were performed using SPSS version 23.0 for Windows.

All imaging was conducted on a 3.0-T Siemens MRI scanner (Erlangen, Germany) at the Shanghai Key Laboratory of Magnetic Resonance. High-resolution T1-weighted structural images were acquired using a magnetization-prepared rapid gradient echo (MPRAGE) sequence with 192 sagittal slices and the following parameters: repetition time (TR) = 1900 ms; echo time (TE) = 3.42 ms; inversion time (TI) = 900 ms; flip angle = 9°; field of view (FOV) = 240 × 240 mm²; matrix = 256 × 256; slice thickness/gap = 1/0.5 mm. T2-weighted images for lesion localization were obtained with a turbo spin-echo sequence (30 axial slices; TR = 6,000 ms; TE = 93 ms; FOV = 220 × 220 mm²; flip angle = 120°; matrix = 320 × 320; slice thickness/gap = 5/0 mm). Rs-fMRI data were collected using an echo-planar imaging (EPI) sequence (30 axial slices; TR = 2000 ms; TE = 30 ms; flip angle = 90°; FOV = 220 × 220 mm²; matrix = 64 × 64; slice thickness/gap = 4/0.8 mm; 240 volumes). The total acquisition time was 8 min 6 s, including a 6-s dummy scan to ensure magnetization equilibrium. During rs-fMRI scanning, participants were instructed to remain still, keep their eyes closed, stay awake, and avoid focused thought.

Lesion volumes were manually delineated on T1- and T2-weighted images by two experienced radiologists using MRIcron software¹. Lesions were outlined slice by slice for each patient, with T1-weighted images used for primary outlining and T2-weighted images used for validation. To standardize lesion laterality, all images were mirrored so that lesions were represented on the left hemisphere (Figure 1). The resulting lesion masks were used for visualization, lesion volume calculation, and exclusion of intra-lesion voxels during dFC analysis. Notably, lesion volume was not significantly correlated with UL motor function, as measured by UL-FMA scores (Pearson’s r = −0.208, p = 0.087), suggesting that stroke size alone did not account for impairment severity in this cohort.

Image preprocessing and analysis were performed using Statistical Parametric Mapping (SPM12²) and the Data Processing Assistant for RS-fMRI (DPABI³). For patients with right-hemisphere lesions, images were flipped along the midsagittal plane to standardize orientation, designating the left hemisphere as ipsilesional and the right as contralesional for all participants. Preprocessing steps were as follows: (1) The first 10 volumes of each functional scan were discarded to allow for magnetization equilibrium and participant adaptation. The remaining 230 volumes were corrected for slice timing, and head motion. No subject was excluded, as none exceeded the motion threshold of 2.0 mm translation or 2.0° rotation; (2) Linear trends, mean white matter and cerebrospinal fluid (CSF) signals, and 24 motion parameters (translations, rotations, their temporal derivatives, and quadratic terms) were regressed out of each voxel time series. Global signal regression was not applied to preserve correlation matrix structure, particularly for sliding-window analyses (32, 33); (3) Functional images were spatially normalized to Montreal Neurological Institute (MNI) space using each patient’s lesion mask and the DARTEL (Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra) algorithm (34). Images were resampled to 3 × 3 × 3 mm³ voxels using the DARTEL-derived deformation fields; (4) Time-series data were spatially smoothed with a 6 mm full-width at half-maximum (FWHM) Gaussian kernel, linearly detrended, and band-pass filtered (0.01–0.1 Hz) to reduce physiological and low-frequency noise.

DFC analysis was conducted using the Temporal Dynamic Analysis (TDA) module within the DPABI toolbox (35). A sliding-window approach was applied to capture temporal fluctuations in connectivity between the left primary M1 and all other brain voxels (36). The left M1 seed was defined at MNI coordinates (−38, −22, 56), consistent with prior studies implicating this region in UL motor function after stroke (3, 12, 16, 37, 38). Each sliding window covered 22 TRs (44 s) and shifted by 1 TR (2 s), enabling fine-grained tracking of temporal dynamics (19). This window length was selected according to the lower limit of the frequency spectrum (<0.5/f_lower = 50 s; here, f_lower = 0.01 Hz), ensuring reliable estimation while preserving sensitivity to short-term fluctuations (39). A 22-TR window therefore represents a balance between temporal resolution and estimation stability, consistent with prior literature (16, 19, 39). From the 230 rs-fMRI volumes, each participant contributed 209 overlapping sliding windows. Within each window, Pearson correlation coefficients were calculated between the left M1 time series and all other brain voxels, generating a dynamic series of whole-brain correlation maps (Figure 2). Fisher’s r-to-z transformation was then applied to improve normality of the correlation values (40).

The 3D volumetric functional connectivity maps were used as inputs to a 3D-CNN to classify participants into either the MSP or SSP group. Consistent with prior neuroimaging studies, no standard data augmentation techniques were applied, as introducing synthetic data could bias the training process. Rs-fMRI signals are particularly sensitive; even a 3 mm spatial shift can substantially alter their interpretation. For this reason, conventional augmentation strategies are generally discouraged in the neuroimaging community (24, 26, 28).

A 3D-CNN–based deep learning framework was developed for participant classification. The model was implemented in MATLAB 2019b with GPU acceleration on an NVIDIA Quadro RTX 5000. Training employed the Adam optimizer (41) with an initial learning rate of 0.001, which was reduced by half every 10 epochs. Each epoch consisted of 32-sample mini-batches over 50 epochs in total. An epsilon value of 0.001 was applied, and cross-entropy loss was used as the cost function. To mitigate overfitting, 10-fold cross-validation was performed to estimate mean classification accuracy at the dFC level. For each fold, 10% of the dFC maps were reserved for testing, while the remaining data were split into training (80%) and validation (10%) sets. The network was based on a modified VGG-Net architecture (42), incorporating batch normalization in convolutional layers and Rectified Linear Unit (ReLU) activation functions. A dropout rate of 0.7 was applied to the fully connected layers to enhance generalization. Weight initialization followed the method proposed by He et al. (43), using a zero-mean Gaussian distribution with a standard deviation scaled to network depth. Hyperparameters—including learning rate, epsilon, dropout rate, batch size, and number of epochs—were optimized based on prior work (24, 26). Specifically, epsilon was tuned in the range [0.1:0.05:1], learning rate across a logarithmic scale [1, 0.1, 0.01, 0.001, 0.0001, 0.00001], and dropout rate between [0.1:0.05:1]. Batch size was optimized using the maximum available GPU memory, while the number of epochs was tuned within [10:1:200]. The complete 3D-CNN architecture is shown in Figure 3 and detailed in Table 1.

To examine the contribution of common functional connectivity to deep learning classification and to better understand how deep neural networks iteratively optimize classification weights (24), we analyzed dFC between the ipsilesional M1 and other brain regions across individual sliding windows. Because multiple pairwise comparisons can introduce statistical bias, we applied a k-means clustering algorithm to dFC maps from all participants to identify distinct connectivity states at rest and to assess intra-individual variability. Following previous studies (44), L1 distance was chosen as the similarity metric due to its robustness in high-dimensional data. The clustering procedure consisted of two main steps. First, exemplar datasets were selected based on local maxima in functional connectivity variance across windows, consistent with prior work (19, 23). Second, rather than relying on conventional criteria such as the elbow method or Silhouette index—which may not guarantee that each participant is represented across multiple states—we determined the optimal number of clusters (k) by evaluating the participation rate (PR) of subjects across candidate solutions (19, 23). Cluster solutions were examined for k values ranging from 2 to 8. A three-cluster solution (k = 3) was selected because it yielded an adequate distribution of subject participation (state 1: PR = 1/69; state 2: PR = 68/69; state 3: PR = 69/69). States 2 and 3 were retained for further analysis, while state 1 was excluded due to insufficient subject representation, which precluded statistically meaningful comparisons.

To evaluate the statistical significance of our results, we performed a permutation test on the 3D-CNN classification accuracies, followed by a two-sample t-test to assess between-group differences. For the classification analysis, test data labels were randomly shuffled 1,000 times within each of the 10 cross-validation folds to estimate the probability of achieving an accuracy greater than that obtained with the true labels. Between-group differences in dFC (MSP vs. SSP) were tested using two-sample t-tests, with age, sex, and duration of illness included as covariates. A significance threshold of uncorrected p < 0.05 was adopted.

No significant differences were observed between the MSP and SSP groups in age (MSP: 54.3 ± 11.95 years; SSP: 51.28 ± 10.73 years; p = 0.392, two-sample t-test), sex distribution (MSP: 32 males; SSP: 24 males; p = 0.513, χ² test), duration of illness (MSP: 8.01 ± 6.5 months; SSP: 9.40 ± 7.10 months; p = 0.410, two-sample t-test), or lesion volume (MSP: 6.7 ± 4.73 mL; SSP: 4.97 ± 4.70 mL; p = 0.513, two-sample t-test). As expected, UL-FMA scores were significantly different, with higher scores in the MSP group compared to the SSP group (31.4 ± 7.84 vs. 8.85 ± 4.49; p < 0.0001). Detailed demographic and clinical characteristics for all participants, including healthy controls (HCs), are presented in Table 2.

The 3D-CNN achieved high classification performance under 10-fold cross-validation. To mitigate potential bias from class imbalance, balanced accuracy was reported alongside other evaluation metrics. The mean performance across folds was as follows: training accuracy = 99.23%, test accuracy = 99.80%, specificity = 99.86%, sensitivity = 99.74%, F-score = 99.78%, and balanced accuracy = 99.80% (Table 3).

Statistical significance was confirmed using permutation testing. A threshold of p = 0.001 was applied across all folds, and permutation tests consistently yielded p < 0.001, indicating that the observed classification performance was highly unlikely to arise by chance. Fold-wise permutation-derived p-values are presented in Table 3.

Discriminative brain regions contributing to the deep learning framework were identified using two-sample t-tests across dynamic states 2 and 3. Significant regions included the angular gyrus (Angular), MOF, SFG, SPL, SMA, superior temporal pole (STP), thalamus, cerebellum, gyrus rectus, middle temporal gyrus (MTG), and precuneus. Uncorrected t-values indicated group differences in functional connectivity: indices A-E corresponded to state 2 (t = 2.64–3.37), and indices F-M corresponded to state 3 (t = 2.88–3.75) (Table 4 and Figure 4). In Figure 4, red shading denotes regions where functional connectivity was higher in the MSP group compared to the SSP group, while blue shading indicates higher connectivity in SSP relative to MSP.