RESUMO
Resting-state functional connectivity (RSFC) has been widely adopted for individualized trait prediction. However, multiple confounding factors may impact the predicted brain-behavior relationships. In this study, we investigated the impact of 4 confounding factors including time series length, functional connectivity (FC) type, brain parcellation choice, and variance of the predicted target. The data from Human Connectome Project including 1,206 healthy subjects were employed, with 3 cognitive traits including fluid intelligence, working memory, and picture vocabulary ability as the prediction targets. We compared the prediction performance under different settings of these 4 factors using partial least square regression. Results demonstrated appropriate time series length (300 time points) and brain parcellation (independent component analysis, ICA100/200) can achieve better prediction performance without too much time consumption. FC calculated by Pearson, Spearman, and Partial correlation achieves higher accuracy and lower time cost than mutual information and coherence. Cognitive traits with larger variance among subjects can be better predicted due to the well elaboration of individual variability. In addition, the beneficial effects of increasing scan duration to prediction partially arise from the improved test-retest reliability of RSFC. Taken together, the study highlights the importance of determining these factors in RSFC-based prediction, which can facilitate standardization of RSFC-based prediction pipelines going forward.
Assuntos
Conectoma , Imageamento por Ressonância Magnética , Humanos , Reprodutibilidade dos Testes , Imageamento por Ressonância Magnética/métodos , Encéfalo , Conectoma/métodos , CogniçãoRESUMO
BACKGROUND: The rupture of intracranial aneurysms (IAs) would result in subarachnoid hemorrhage with high mortality and disability. Predicting the risk of IAs rupture remains a challenge. METHODS: This paper proposed an effective method for classifying IAs rupture status by integrating a PointNet-based model and machine learning algorithms. First, medical image segmentation and reconstruction algorithms were applied to 3D Digital Subtraction Angiography (DSA) imaging data to construct three-dimensional IAs geometric models. Geometrical parameters of IAs were then acquired using Geomagic, followed by the computation of hemodynamic clouds and hemodynamic parameters using Computational Fluid Dynamics (CFD). A PointNet-based model was developed to extract different dimensional hemodynamic cloud features. Finally, five types of machine learning algorithms were applied on geometrical parameters, hemodynamic parameters, and hemodynamic cloud features to classify and recognize IAs rupture status. The classification performance of different dimensional hemodynamic cloud features was also compared. RESULTS: The 16-, 32-, 64-, and 1024-dimensional hemodynamic cloud features were extracted with the PointNet-based model, respectively, and the four types of cloud features in combination with the geometrical parameters and hemodynamic parameters were respectively applied to classify the rupture status of IAs. The best classification outcomes were achieved in the case of 16-dimensional hemodynamic cloud features, the accuracy of XGBoost, CatBoost, SVM, LightGBM, and LR algorithms was 0.887, 0.857, 0.854, 0.857, and 0.908, respectively, and the AUCs were 0.917, 0.934, 0.946, 0.920, and 0.944. In contrast, when only utilizing geometrical parameters and hemodynamic parameters, the accuracies were 0.836, 0.816, 0.826, 0.832, and 0.885, respectively, with AUC values of 0.908, 0.922, 0.930, 0.884, and 0.921. CONCLUSION: In this paper, classification models for IAs rupture status were constructed by integrating a PointNet-based model and machine learning algorithms. Experiments demonstrated that hemodynamic cloud features had a certain contribution weight to the classification of IAs rupture status. When 16-dimensional hemodynamic cloud features were added to the morphological and hemodynamic features, the models achieved the highest classification accuracies and AUCs. Our models and algorithms would provide valuable insights for the clinical diagnosis and treatment of IAs.
RESUMO
Background: Acute hematologic toxicity (HT) is a prevalent adverse tissue reaction observed in cervical cancer patients undergoing chemoradiotherapy (CRT), which may lead to various negative effects such as compromised therapeutic efficacy and prolonged treatment duration. Accurate prediction of HT occurrence prior to CRT remains challenging. Methods: A discovery dataset comprising 478 continuous cervical cancer patients (140 HT patients) and a validation dataset consisting of 205 patients (52 HT patients) were retrospectively enrolled. Both datasets were categorized into the CRT group and radiotherapy (RT)-alone group based on the treatment regimen, i.e., whether chemotherapy was administered within the focused RT duration. Radiomics features were derived by contouring three regions of interest (ROIs)-bone marrow (BM), femoral head (FH), and clinical target volume (CTV)-on the treatment planning CT images before RT. A comprehensive model combining the radiomics features as well as the demographic, clinical, and dosimetric features was constructed to classify patients exhibiting acute HT symptoms in the CRT group, RT group, and combination group. Furthermore, the time-to-event analysis of the discriminative ROI was performed on all patients with acute HT to understand the HT temporal progression. Results: Among three ROIs, BM exhibited the best performance in classifying acute HT, which was verified across all patient groups in both discovery and validation datasets. Among different patient groups in the discovery dataset, acute HT was more precisely predicted in the CRT group [area under the curve (AUC) = 0.779, 95% CI: 0.657-0.874] than that in the RT-alone (AUC = 0.686, 95% CI: 0.529-0.817) or combination group (AUC = 0.748, 95% CI: 0.655-0.827). The predictive results in the validation dataset similarly coincided with those in the discovery dataset: CRT group (AUC = 0.802, 95% CI: 0.669-0.914), RT-alone group (AUC = 0.737, 95% CI: 0.612-0.862), and combination group (AUC = 0.793, 95% CI: 0.713-0.874). In addition, distinct feature sets were adopted for different patient groups. Moreover, the predicted HT risk of BM was also indicative of the HT temporal progression. Conclusions: HT prediction in cervical patients is dependent on both the treatment regimen and ROI selection, and BM is closely related to the occurrence and progression of HT, especially for CRT patients.
RESUMO
It is of potential clinical value to improve the accuracy of Alzheimer's disease (AD) recognition using structural MRI. We proposed a reparametrized convolutional neural network (Re-CNN) to discriminate AD from NC by applying morphological metrics and deep semantic features. The deep semantic features were extracted through Re-CNN on structural MRI. Considering the high redundancy in deep semantic features, we constrained the similarity of the features and retained the most distinguishing features utilizing the reparametrized module. The Re-CNN model was trained in an end-to-end manner on structural MRI from the ADNI dataset and tested on structural MRI from the AIBL dataset. Our proposed model achieves better performance over some existing structural MRI-based AD recognition models. The experimental results show that morphological metrics along with the constrained deep semantic features can relatively improve AD recognition performance. Our code is available at: https://github.com/czp19940707/Re-CNN.
RESUMO
Background: The rupture risk assessment of intracranial aneurysms (IAs) is clinically relevant. How to accurately assess the rupture risk of IAs remains a challenge in clinical decision-making. Purpose: We aim to build an integrated model to improve the assessment of the rupture risk of IAs. Materials and Methods: A total of 148 (39 ruptured and 109 unruptured) IA subjects were retrospectively computed with computational fluid dynamics (CFDs), and the integrated models were proposed by combining machine learning (ML) and deep learning (DL) algorithms. ML algorithms that include random forest (RF), k-nearest neighbor (KNN), XGBoost (XGB), support vector machine (SVM), and LightGBM were, respectively, adopted to classify ruptured and unruptured IAs. A Pointnet DL algorithm was applied to extract hemodynamic cloud features from the hemodynamic clouds obtained from CFD. Morphological variables and hemodynamic parameters along with the extracted hemodynamic cloud features were acted as the inputs to the classification models. The classification results with and without hemodynamic cloud features are computed and compared. Results: Without consideration of hemodynamic cloud features, the classification accuracy of RF, KNN, XGB, SVM, and LightGBM was 0.824, 0.759, 0.839, 0.860, and 0.829, respectively, and the AUCs of them were 0.897, 0.584, 0.892, 0.925, and 0.890, respectively. With the consideration of hemodynamic cloud features, the accuracy successively increased to 0.908, 0.873, 0.900, 0.926, and 0.917. Meanwhile, the AUCs reached 0.952, 0.881, 0.950, 0.969, and 0.965 eventually. Adding consideration of hemodynamic cloud features, the SVM could perform best with the highest accuracy of 0.926 and AUC of 0.969, respectively. Conclusion: The integrated model combining ML and DL algorithms could improve the classification of IAs. Adding consideration of hemodynamic cloud features could bring more accurate classification, and hemodynamic cloud features were important for the discrimination of ruptured IAs.