This study aims to explore the value of a machine learning (ML) model based on radiomics features and clinical features in predicting the outcome of spontaneous supratentorial intracerebral hemorrhage (sICH) 90 days after surgery. A total of 348 patients with sICH underwent craniotomy evacuation of hematoma from three medical centers. One hundred and eight radiomics features were extracted from sICH lesions on baseline CT. Radiomics features were screened using 12 feature selection algorithms. Clinical features included age, gender, admission Glasgow Coma Scale (GCS), intraventricular hemorrhage (IVH), midline shift (MLS), and deep ICH. Nine ML models were constructed based on clinical feature, and clinical features + radiomics features, respectively. Grid search was performed on different combinations of feature selection and ML model for parameter tuning. The averaged receiver operating characteristics (ROC) area under curve (AUC) was calculated and the model with the largest AUC was selected. It was then tested using multicenter data. The combination of lasso regression feature selection and logistic regression model based on clinical features + radiomics features had the best performance (AUC: 0.87). The best model predicted an AUC of 0.85 (95%CI, 0.75–0.94) on the internal test set and 0.81 (95%CI, 0.64–0.99) and 0.83 (95%CI, 0.68–0.97) on the two external test sets, respectively. Twenty-two radiomics features were selected by lasso regression. The second-order feature gray level non-uniformity normalized was the most important radiomics feature. Age is the feature with the greatest contribution to prediction. The combination of clinical features and radiomics features using logistic regression models can improve the outcome prediction of patients with sICH 90 days after surgery. Full article

Altered cerebral structure and function have been observed in young survivors of acute lymphoblastic leukemia (ALL). However, the topological organization of the morphological brain networks (MBNs) has not yet been investigated at the individual level. Twenty-three young survivors of ALL and twenty healthy controls (HCs) were recruited and underwent T1-weighted magnetic resonance imaging (MRI) scanning. After preprocessing and segmentation, individual-based MBNs were constructed based on the morphological similarity of gray matter using the combined Euclidean distance. Young survivors showed a significantly lower global clustering coefficient (p = 0.008) and local efficiency (p = 0.035) compared with HCs. In addition, ALL survivors exhibited bidirectional alterations (decreases and increases) in nodal centrality and efficiency around the Rolandic operculum and posterior occipital lobe (p < 0.05, false discovery rate (FDR) corrected). Altered nodal topological efficiencies were associated with off-therapy duration and verbal memory capacity in the digit span test (p < 0.05, FDR corrected). Network-based statistical analysis revealed decreased morphological connections mainly in the pallidum subnetwork, which was negatively correlated with off-therapy durations (p < 0.05). Overall, the topological organization of the individual-based MBNs was disrupted in the young survivors of ALL, which may play a crucial role in executive efficiency deficits. Full article

Background and purpose: Previous studies have shown that Hypoperfusion Intensity Ratio (HIR) derived from Perfusion Imaging (PWI) associated with collateral status in large-vessel occlusion (LVO) acute ischemic stroke (AIS) and could predict the rate of collateral flow, speed of infarct growth, and clinical outcome after endovascular treatment (EVT). We hypothesized that HIR derived from CT Perfusion (CTP) imaging could relatively accurately predict the functional outcome in LVO AIS patients receiving different types of treatment. Methods: Imaging and clinical data of consecutive patients with LVO AIS were retrospectively reviewed. Multi-phase CT angiography (mCTA) scoring was performed by 2 blinded neuroradiologists. CTP images were processed using an automatic post-processing analysis software. Correlation between the HIR and the functional outcome was calculated using the Spearman correlation. The efficacy of the HIR and the CTA collateral scores for predicting prognosis were compared. The optimal threshold of the HIR for predicting favorable functional outcome was determined using receiver operating characteristic (ROC) curve analysis. Results: 235 patients with LVO AIS were included. Patients with favorable functional outcome had lower HIR (0.1 [interquartile range (IQR), 0.1–0.2]) vs. 0.4 (IQR, 0.4–0.5)) and higher mCTA collateral scores (3 [IQR, 3–4] vs. 3 [IQR, 2–3]; p < 0.001) along with smaller infarct core volume (2.1 [IQR, 1.0–4.5]) vs. (15.2 [IQR, 5.5–39.3]; p < 0.001), larger mismatch ratio (22.9 [IQR, 11.6–45.6]) vs. (5.8 [IQR, 2.6–14]); p < 0.001), smaller ischemic volume (59.0 [IQR, 29.7–89.2]) vs. (97.5 [IQR, 68.7–142.2]; p < 0.001), and smaller final infarct volume (12.6 [IQR, 7.5–18.4]) vs. (78.9 [IQR, 44.5–165.0]; p < 0.001) than those with unfavorable functional outcome. The HIR was significantly positively correlated with the functional outcome [r = 0.852; 95% confidence interval (CI): 0.813–0.884; p < 0.0001]. The receiver operating characteristic (ROC) analysis showed that the optimal threshold for predicting a favorable functional outcome was HIR ≤ 0.3 [area under the curve (AUC) 0.968; sensitivity 88.89%; specificity 99.21%], which was higher than the mCTA collateral score [AUC 0.741; sensitivity 82.4%; specificity 48.8%]. Conclusions: HIR was associated with the functional outcome of LVO AIS patients, and the correlation coefficient was higher than mCTA collateral score. HIR outperformed mCTA collateral score in predicting functional outcome. Full article

Background: Deep learning (DL) could predict isocitrate dehydrogenase (IDH) mutation status from MRIs. Yet, previous work focused on CNNs with refined tumor segmentation. To bridge the gap, this study aimed to evaluate the feasibility of developing a Transformer-based network to predict the IDH mutation status free of refined tumor segmentation. Methods: A total of 493 glioma patients were recruited from two independent institutions for model development (TCIA; N = 259) and external test (AHXZ; N = 234). IDH mutation status was predicted directly from T2 images with a Swin Transformer and conventional ResNet. Furthermore, to investigate the necessity of refined tumor segmentation, seven strategies for the model input image were explored: (i) whole tumor slice; (ii–iii) tumor mask and/or not edema; (iv–vii) tumor bounding box of 0.8, 1.0, 1.2, 1.5 times. Performance comparison was made among the networks of different architectures along with different image input strategies, using area under the curve (AUC) and accuracy (ACC). Finally, to further boost the performance, a hybrid model was built by incorporating the images with clinical features. Results: With the seven proposed input strategies, seven Swin Transformer models and seven ResNet models were built, respectively. Based on the seven Swin Transformer models, an averaged AUC of 0.965 (internal test) and 0.842 (external test) were achieved, outperforming 0.922 and 0.805 resulting from the seven ResNet models, respectively. When a bounding box of 1.0 times was used, Swin Transformer (AUC = 0.868, ACC = 80.7%), achieved the best results against the one that used tumor segmentation (Tumor + Edema, AUC = 0.862, ACC = 78.5%). The hybrid model that integrated age and location features into images yielded improved performance (AUC = 0.878, Accuracy = 82.0%) over the model that used images only. Conclusions: Swin Transformer outperforms the CNN-based ResNet in IDH prediction. Using bounding box input images benefits the DL networks in IDH prediction and makes the IDH prediction free of refined glioma segmentation feasible. Full article

Purpose: To develop a preliminary deep learning model that uses diffusion-weighted imaging (DWI) images to classify the severity of neurological impairment caused by ischemic stroke. Materials and Methods: This retrospective study included 851 ischemic stroke patients (711 patients in the training set and 140 patients in the test set). The patients’ NIHSS scores, which reflect the severity of neurological impairment, were reviewed upon admission and on Day 7 of hospitalization and were classified into two stages (stage 1 for NIHSS < 5 and stage 2 for NIHSS ≥ 5). A 3D-CNN was trained to predict the stage of NIHSS based on different preprocessed DWI images. The performance in predicting the severity of anterior and posterior circulation stroke was also investigated. The AUC, specificity, and sensitivity were calculated to evaluate the performance of the model. Results: Our proposed model obtained better performance in predicting the NIHSS stage on Day 7 of hospitalization than that at admission (best AUC 0.895 vs. 0.846). Model D trained with DWI images (normalized with z-score and resized to 256 × 256 × 64 voxels) achieved the best AUC of 0.846 in predicting the NIHSS stage at admission. Model E rained with DWI images (normalized with maximum–minimum and resized to 128 × 128 × 32 voxels) achieved the best AUC of 0.895 in predicting the NIHSS stage on Day 7 of hospitalization. Our model also showed promising performance in predicting the NIHSS stage on Day 7 of hospitalization for anterior and posterior circulation stroke, with the best AUCs of 0.905 and 0.903, respectively. Conclusions: Our proposed 3D-CNN model can effectively predict the neurological severity of IS using DWI images and performs better in predicting the NIHSS stage on Day 7 of hospitalization. The model also obtained promising performance in subgroup analysis, which can potentially help clinical decision making. Full article

Background: Multiple lesions are uncommon in brain gliomas, and their pathophysiology is poorly understood. Invasive growth along white matter tracts is an important clinicopathological characteristic of gliomas, and a major factor in a poor therapeutic outcome. Here, we used probabilistic fiber tracking and cluster analysis to investigate the inter-focal connectivity relationships of multiple gliomas, in order to seek inferential evidence of common origin. Methods: MRI scans of 46 patients with multiple gliomas were retrospectively analyzed. Before surgery, all patients underwent multimodal functional MR imaging, including diffusion tensor imaging, enhanced 3D T1-weighted imaging, diffusion-weighted imaging, 1H MR spectroscopy, and dynamic susceptibility contrast perfusion-weighted imaging. Probabilistic fiber tracking was used to quantify white matter connectivity between neoplastic foci. Hierarchical cluster analysis was performed to identify patterns of white matter connection. Results: Cluster analysis reveals two patterns of connectivity, one with smaller, and one with greater, connectivity (2675 ± 1098 versus 30432 ± 22707, p < 0.0001). The two subgroups show significant differences in relative cerebral blood volume (2.31 ± 0.95 versus 1.73 ± 0.48, p = 0.002) and lipid/creatine ratio (0.32 ± 0.22 versus 0.060 ± 0.051, p = 0.006). Conclusion: Two distinct patterns of white matter connection exist in multiple gliomas. Those with lower connectivity tend to have independent origins, and can be termed true multicentric glioma, whereas those with greater connectivity tend to share common origin, and spread along white matter tracts. True multicentric gliomas have higher vascularity and more intratumoral necrosis. These findings may help to develop personalized therapeutic strategies for multiple gliomas. Full article

This study aimed to investigate the feasibility of predicting oxygen 6-methylguanine-DNA methyltransferase (MGMT) promoter methylation in diffuse gliomas by developing a deep learning approach using MRI radiomics. A total of 111 patients with diffuse gliomas participated in the retrospective study (56 patients with MGMT promoter methylation and 55 patients with MGMT promoter unmethylation). The radiomics features of the two regions of interest (ROI) (the whole tumor area and the tumor core area) for four sequences, including T1 weighted image (T1WI), T2 weighted image (T2WI), apparent diffusion coefficient (ADC) maps, and T1 contrast-enhanced (T1CE) MR images were extracted and jointly fed into the residual network. Then the deep learning method was developed and evaluated with a five-fold cross-validation, where in each fold, the dataset was randomly divided into training (80%) and validation (20%) cohorts. We compared the performance of all models using area under the curve (AUC) and average accuracy of validation cohorts and calculated the 10 most important features of the best model via a class activation map. Based on the ROI of the whole tumor, the predictive capacity of the T1CE and ADC model achieved the highest AUC value of 0.85. Based on the ROI of the tumor core, the T1CE and ADC model achieved the highest AUC value of 0.90. After comparison, the T1CE combined with the ADC model based on the ROI of the tumor core exhibited the best performance, with the highest average accuracy (0.91) and AUC (0.90) among all models. The deep learning method using MRI radiomics has excellent diagnostic performance with a high accuracy in predicting MGMT promoter methylation in diffuse gliomas. Full article

In-stent restenosis (ISR) after carotid artery stenting (CAS) critically influences long-term CAS benefits and safety. The study was aimed at screening preoperative ISR-predictive features and developing predictive models. Thus, we retrospectively analyzed clinical and imaging data of 221 patients who underwent pre-CAS carotid computed tomography angiography (CTA) and whose digital subtraction angiography data for verifying ISR presence were available. Carotid plaque characteristics determined using CTA were used to build a traditional model. Backward elimination (likelihood ratio) was used for the radiomics model. Furthermore, a combined model was built using the traditional and radiomics features. Five-fold cross-validation was used to evaluate the accuracy of the trained classifier and stability of the selected features. Follow-up angiography showed ISR in 30 patients. Carotid plaque length and thickness were independently associated with ISR (multivariate analysis); regarding the conventional model, the area under the curve (AUC) was 0.84 and 0.82 in the training and validation cohorts, respectively. The corresponding AUC values for the radiomics-based model were 0.87 and 0.82, and those for the optimal combined model were 0.88 and 0.83. Plaque length and thickness could independently predict post-CAS ISR, and the combination of radiomics and plaque features afforded the best predictive performance. Full article

Glioma grading plays an important role in surgical resection. We investigated the ability of different feature reduction methods in support vector machine (SVM)-based diffusion kurtosis imaging (DKI) histogram parameters to distinguish glioma grades. A total of 161 glioma patients who underwent magnetic resonance imaging (MRI) from January 2017 to January 2020 were included retrospectively. The patients were divided into low-grade (n = 61) and high-grade (n = 100) groups. Parametric DKI maps were derived, and 45 features from the DKI maps were extracted semi-automatically for analysis. Three feature selection methods [principal component analysis (PCA), recursive feature elimination (RFE) and least absolute shrinkage and selection operator (LASSO)] were used to establish the glioma grading model with an SVM classifier. To evaluate the performance of SVM models, the receiver operating characteristic (ROC) curves of SVM models for distinguishing glioma grades were compared with those of conventional statistical methods. The conventional ROC analysis showed that mean diffusivity (MD) variance, MD skewness and mean kurtosis (MK) C50 could effectively distinguish glioma grades, particularly MD variance. The highest classification distinguishing AUC was found using LASSO at 0.904 ± 0.069. In comparison, classification AUC by PCA was 0.866 ± 0.061, and 0.899 ± 0.079 by RFE. The SVM-PCA model with the lowest AUC among the SVM models was significantly better than the conventional ROC analysis (z = 1.947, p = 0.013). These findings demonstrate the superiority of DKI histogram parameters by LASSO analysis and SVM for distinguishing glioma grades. Full article

The aim of this study was combining multi-level resting-state functional magnetic resonance imaging (rs-fMRI) features with machine learning method to distinguish breast cancer patients with chemotherapy-related subjective cognitive complaints (SCC) from non-chemotherapy (BC) and healthy controls (HC). Forty subjects in SCC group, forty-nine in BC group and thirty-four in HC group were recruited and underwent rs-fMRI scanning. Based on the anatomical automatic labeling brain atlas, the functional metrics of all subjects included functional connectivity, amplitude of low frequency fluctuation and fractional amplitude of low frequency fluctuation, regional homogeneity, voxel-mirrored homotopic connectivity and degree centrality were calculated and extracted as features set. Then, the rs-fMRI features were selected by two-sample t-test, removing variables with a high pairwise correlation and least absolute shrinkage and selection operator regression. Finally, the support vector machine models were built for classification (SCC vs. BC, SCC vs. HC). Thirty-eight features (SCC vs. BC) and seventeen features (SCC vs. HC) were selected separately, and the accuracy of the models were 82.0% and 91.9%, respectively. These findings demonstrated a valid machine learning approach that effectively distinguished breast cancer patients with chemotherapy-related SCC from non-chemotherapy and healthy controls, providing potential neuroimaging evidence for early diagnosis and clinical intervention of chemotherapy-related SCC. Full article

This study aimed to delineate cortico-striato-thalamo-cerebellar network profiles based on static and dynamic connectivity analysis in genetic generalized and focal epilepsies with generalized tonic-clonic seizures, and to evaluate its potential for distinguishing these two epilepsy syndromes. A total of 342 individuals participated in the study (114 patients with genetic generalized epilepsy with generalized tonic-clonic seizures (GE-GTCS), and 114 age- and sex-matched patients with focal epilepsy with focal to bilateral tonic-clonic seizure (FE-FBTS), 114 healthy controls). Resting-state fMRI data were examined through static and dynamic functional connectivity (dFC) analyses, constructing cortico-striato-thalamo-cerebellar networks. Network patterns were compared between groups, and were correlated to epilepsy duration. A pattern-learning algorithm was applied to network features for classifying both epilepsy syndromes. FE-FBTS and GE-GTCS both presented with altered functional connectivity in subregions of the motor/premotor and somatosensory networks. Among these two groups, the connectivity within the cerebellum increased in the static, while the dFC variability decreased; conversely, the connectivity of the thalamus decreased in FE-FBTS and increased in GE-GTCS in the static state. Connectivity differences between patient groups were mainly located in the thalamus and cerebellum, and correlated with epilepsy duration. Support vector machine (SVM) classification had accuracies of 66.67%, 68.42%, and 77.19% when using static, dynamic, and combined approaches to categorize GE-GTCS and FE-GTCS. Network features with high discriminative ability predominated in the thalamic and cerebellar connectivities. The network embedding of the thalamus and cerebellum likely plays an important differential role in GE-GTCS and FE-FBTS, and could serve as an imaging biomarker for differential diagnosis. Full article