3097
Radiomics Model for Prognosis of Brainstem Stroke Based on Lesion and Surrounding Features1Harbin Medical University Second Affiliated Hospital, Harbin, China, 2GE Healthcare, MR Research China, Beijing, China
Synopsis
Keywords: Radiomics, Radiomics, Stroke
Motivation: Stroke is a major global health issue, necessitating early outcome prediction for optimal treatment. Brainstem stroke, often overlooked, requires dedicated predictive models due to its unique challenges.
Goal(s): Develop radiomics models to predict brainstem stroke outcomes, considering infarct edge and surrounding regions, improving prognosis, and simplifying clinical evaluation.
Approach: 474 patients were studied, and radiomics features were extracted from diffusion-weighted images. Machine learning models were trained using SVM, RF, KNN and AdaBoost algorithms.
Results: The RF model, based on the circle2 region, exhibited the highest performance (AUC=0.84). Models in the circle region outperformed core.
Impact: Our specialized radiomics models offer a valuable tool for personalized brainstem stroke treatment planning, potentially enhancing patient outcomes.
Introduction
Stroke is global health concern, ranking among the leading causes of death and disability worldwide[1]. Unfortunately, due to the narrow treatment time windows, many stroke patients can only receive treatments such as antiplatelet therapy and neuroprotective agents. Early assessment of treatment outcomes can significantly benefit more patients. Recent studies have indicated that radiomic features extracted from magnetic resonance imaging (MRI) have the potential to predict stroke outcomes. However, brainstem stroke lesions are relatively small compared to those in the cerebrum and cerebellum, making them less amenable to traditional radiomic analysis. Additionally, patients with brainstem stroke often face a more challenging prognosis. Therefore, the development of a specialized radiomic prediction model for brainstem ischemic stroke is imperative.Methods
In this study, we enrolled a cohort of 474 patients with brainstem stroke with approvement from the institutional review board. The prognosis was assessed based on the modified Rankin Scale (MRS) at 6 months after discharge, categorizing it as either good (MRS ≤ 2) or poor (MRS > 2). Subsequently, the patients were randomly divided into training and testing cohorts with an 8:2 ratio. We expanded the lesion edge outward by 1 and 2 pixels, yielding the dilation1, dilation2, periphery1 and periphery2 regions of the infarct. Furthermore, by eroding the edge of the lesion inward by 1 and 2 pixels, we obtained the core1, core2, rim1 and rim2 regions of the infarct. Additionally, we created two circular regions, circle1 and circle2, by fusing rim1 with periphery1 and rim2 with periphery2, respectively. From each region, we extracted a total of 1744 radiomics features from the diffusion-weighted images (DWI). In order to select the most significant features, we employed the U-test and the least absolute shrinkage and selection operator (LASSO) method. Machine learning models based on the Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (KNN) and AdaBoost algorithms were developed to predict the outcome of conventional treatment. Model performance was evaluated using receiver operating characteristic (ROC) analysis and decision curve analysis. The DeLong test is used to compare statistical differences between different models.Results
After feature extraction and selection, we established a total of four machine learning models to predict outcomes in brainstem stroke patients. Among these models, the RF model based on the circle2 region and the SVM model based on the circle1 region demonstrated superior performance, achieving an impressive area under the ROC curve (AUC) of 0.84 (95% confidence interval, 0.70–0.97) and 0.82 (95% confidence interval, 0.72–0.91) in the testing set, with no significant difference. Decision curve analysis further highlighted the clinical utility of this approach. Except for the core2 region (AUC=0.69) where the KNN model slightly outperformed the circle2 region (AUC=0.68), the performance of all models in the circle1 and circle2 regions was better than the core1 and core2 regions. Specifically, the RF models in the circle2 (AUC=0.84) and core 2 (AUC=0.60) regions showed significant statistical differences (P<0.01). Among the 10 RF models, except for the core1 region, which performed marginally better than the core2 region, the dilation2, periphery2, rim2, and circle2 regions consistently exhibited higher performance compared to the dilation1, periphery1, rim1, and circle1 regions, which have no statistical differences.Discussion
Typically, an infarct consists of a core of ischemia surrounded by ischemic penumbra, often with or without obvious edema[2]. In this study, the circle2 region, which combines the erosion area within the lesion boundary and the expansion area around the edge, demonstrated the most robust performance. This might be attributed to its incorporation of additional features that are not visually discernible but can be analyzed through radiomics techniques. However, in the infarct core area, due to the lack of such characteristics, its performance in our model is inferior compared to other regions. Moreover, dilation2, periphery2, rim2, and circle2 regions exhibited larger areas compared to dilation1, periphery1, rim1, and circle1, potentially encompassing more imaging features, which could explain their enhanced performance. Given the limited volume of the brainstem, where many lesions are only a few pixels in size, establishing and generalizing radiomics models becomes challenging. Therefore, the development of specialized radiomics predictive models specifically tailored to brainstem infarction is crucial. The DWI model we established achieved an AUC value of 0.84, slightly surpassing Guo's predictive model[3] (AUC=0.83), and approaching Zhou's clinical-imaging hybrid model[4](AUC=0.89). This demonstrates excellent predictive performance while simplifying the clinical evaluation process.Conclusion
Radiomics models based on the infarct edge and its surrounding areas offer effective predictions of outcomes in brainstem stroke cases. These models can be invaluable in tailoring personalized treatment plans and ultimately improving patient prognosis.Acknowledgements
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.References
1 Gowda SN, Munakomi S, De Jesus O. Brainstem Stroke. 2023 Jun 24. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2023 Jan.
2 Mandalaneni K, Rayi A, Jillella DV. Stroke Reperfusion Injury. 2022 Oct 31. In: StatPearls [Internet]. Treasure Island (FL): Stat Pearls Publishing; 2023 Jan.
3 Guo Y, Yang Y, Cao F, et al. A Focus on the Role of DSC-PWI Dynamic Radiomics Features in Diagnosis and Outcome Prediction of Ischemic Stroke. J Clin Med. 2022 Sep 13;11(18):5364.
4 Zhou Y, Wu D, Yan S et al. Feasibility of a Clinical-Radiomics Model to Predict the Outcomes of Acute Ischemic Stroke. Korean J Radiol. 2022 Aug;23(8):811-820.
Figures
