Prediction Model for Chronic Intracranial Arterial Occlusion With Radiomic Features

Overview

Chronic intracranial arterial occlusion is associated with a "bidirectional stroke risk," with a significantly increased risk of both ischemic stroke and cerebral hemorrhage. Currently, Western CTAP products, in combination with clinical expertise, offer some predictive value for assessing the risk of ischemic events by evaluating compensatory pathways and overall perfusion in chronic intracranial arterial occlusion. However, there is limited support for assessing the risk of hemorrhagic events. Our proposed project aims to address a significant scientific challenge: the precise assessment of long-term stroke risk in asymptomatic patients with chronic intracranial arterial occlusion using a machine learning-based approach. The rapidly advancing field of machine learning provides a rich set of solutions for tackling this problem. In this project, we intend to develop a deep learning-based segmentation model for key brain regions using multimodal CT scans. Subsequently, we will automate the extraction of radiomic features and CT perfusion parameters, followed by the application of machine learning techniques to construct a stroke risk prediction model tailored for patients with chronic intracranial arterial occlusion.

1. Establishment of a Brain Segmentation Model for Key Brain Regions Based on Multimodal CT The establishment of a brain segmentation model for key brain regions based on multimodal CT is one of the core components of this research. The project plans to use advanced algorithms such as U-Net and Transformer as the foundation to develop a model for automatically segmenting key brain regions in CT images. These key brain regions include the basal ganglia, frontal lobe, temporal lobe, parietal lobe, and others. The development of the automatic segmentation model consists of two main parts: training and validation. Retrospective cohort data from this research are used to constitute the training set. Specific brain regions are manually annotated to serve as training materials for the model. Parameter tuning and internal validation are conducted to achieve satisfactory training results. Prospective cohort data from this research form an independent validation set used for external validation of the model. The performance evaluation criteria for the model include the DICE coefficient and loss value. 2. Exploration of Core Neuroimaging Features Related to Stroke in Patients with Chronic Intracranial Arterial Occlusion Using masks generated by the automatic segmentation model, this study intends to extract radiomic information and CT perfusion parameters of specific brain regions through a self-designed automated workflow. These features will be used to construct a machine learning predictive model for stroke risk in patients with chronic intracranial arterial occlusion. To select the most effective features for stroke prediction, a feature selection and dimensionality reduction process is required. This study plans to use L1 regularization to reduce the dimensionality of the standardized data, aiming to optimize the ROC area under the curve (AUC). Thus, it is possible to explore and refine core neuroimaging features related to stroke in patients with intracranial arterial occlusion, analyze their correlation with stroke, and use them for subsequent model construction. 3. Construction and Validation of a Stroke Prediction Model Based on Core Neuroimaging Features of Key Brain Regions The construction and validation of the prediction model are divided into training and validation phases. The training data are collected from the training cohort and internally validated using cross-validation. The prospective cohort remains an independent validation set for external validation of the model. Model construction begins with testing basic machine learning models' performance. By analyzing the performance of these basic models, the architecture of an ensemble learning model is designed and constructed. Finally, the ensemble learning model is validated. The primary performance metric for the model is the area under the ROC curve (AUC), and additional evaluation metrics such as accuracy, recall, precision, and F1 score are used for supplementary assessment of performance.