DeepStrokeGuard: Explainable and Deployable Deep Learning for Brain Stroke Detection with Class Imbalance Handling
Year: 2026
Type: conference-paper
Abstract: Detection of brain strokes using CT scan is still a difficult task because the class imbalances in normal and stroke rates are high, the false negative rates are also very high and the models lack interpretability. A full framework called DeepStrokeGuard is presented in this paper that reviews various deep learning models (DenseNet121, ResNet50 and Xception) and solves these issues by applying systematic preprocessing, data augmentation, dynamically weighted classes, and threshold tuning using Youden’s J statistic. The fine-tuned ResNet50 has achieved a test accuracy of 99.77%, F1-score of 1.00, AUC of 1.00, precision of 99.5%, and recall of 100% on a test set of 2515 brain CTs. DenseNet121 class weighted and threshold tuned offers a lightweight alternative at 97.25% with 6.96M parameters only. We combine Grad-CAM and LIME to offer visual/pixel evaluation of the extensions of the stroke, and our ablation experiment discloses an enhancement of the accuracy between 68–78% to 99.77%. A web-based framework facilitates real-time CT analysis with confidence scores and thus, the suggested framework can be both very accurate and practically applicable as grounded towards the detection of an early brain stroke.
Institution: Daffodil International University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204488714