Ensemble CNN Model for Accurate Stroke Detection

• Shilpa Bajaj,
• Manju Bala,
• Mohit Angurala,

Vol. 19, No. 12, December 31, 2025
10.3837/tiis.2025.12.006, Download Paper (Free):

• Brain Stroke Detection
• Ensemble CNN
• VGG19
• Machine Learning
• ResNet

Abstract

Timely and accurate diagnosis of stroke plays a vital role in improving patient outcomes. This paper evaluates the performance of a novel ensemble convolutional neural network (CNN) model for stroke prediction using brain CT images. Additionally, two alternative ensemble models are evaluated to compare their effectiveness. Each model applies a different ensemble strategy. The first employs a general CNN-based architecture, incorporating variants such as ResNet and VGG19, where each contributes unique representational strengths. In contrast, the proposed ensemble model combines three pre-trained CNN architectures—ResNet-101, VGG19, and InceptionV2—through transfer learning and fine-tuning. With this approach, the ensemble model improves abstraction of spatial features from complex medical images, while reducing overfitting on smaller datasets. Then the ensemble framework combines the predictive outputs of each individual model and combines them to create a more generalizable and robust classifier than what would be possible with a single model. One of the main innovations of the ensemble model is the performance-weighted prediction approach, which can guide the ensemble to assign quantitative priority to the models with the best classification performance, instead of the conventional static weighting of a typical ensemble. The experimental results also determine that the average ensemble CNN achieved a classification accuracy of 92.43% and an AUC of 0.92. By comparison, the proposed performance-weighted ensemble model achieved an accuracy of 98% and an AUC of 1.0. To further compare the robustness of the proposed model, four models (Ensemble CNN, OzNet-mRMR-NB, Logistic Regression, and the Proposed Ensemble Model) are compared on selected performance metrics. Overall, the results show evidence that performance-weighted ensembles can model the complex medical imaging data and significantly improve classification performance. The results of this research model have not yet been implemented in an operational setting, but they demonstrate potential in augmenting radiologist decision-making and improving diagnostic quality and accuracy in research settings.

Statistics

Cite this article