Scholarlink

<p>Brain stroke occurs due to blockage or rupture in the cerebral blood supply and represents a critical medical emergency requiring rapid and accurate diagnosis. However, manual interpretation of CT scans is time-consuming and may delay clinical decision-making. To address this challenge, this study proposes the Deep Neural Brain Stroke Detection (DNBSD) system, a lightweight deep learning–based framework for automated stroke detection from CT images. The proposed model employs a task-specific convolutional neural network (CNN) architecture consisting of Conv2D, MaxPooling, Batch Normalization, Flatten and Dense layers, containing only 1.67 million trainable parameters and a computational complexity of 0.2973 GFLOPs, making it suitable for resource-constrained clinical environments. Additionally, preprocessing techniques including image resizing and normalization were applied to optimize performance. The model is trained and evaluated on two publicly available datasets: Brain Stroke CT Image Dataset (BSCI) and Brain Stroke Prediction CT Scan Image Dataset (BSPCSI), each divided into training, validation, and testing subsets. Experimental results demonstrate that the DNBSD system achieves high performance, with an accuracy of <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(99.20\%\)</EquationSource></InlineEquation> and an AUC of <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(99.96\%\)</EquationSource></InlineEquation> on the BSCI dataset, and an accuracy of <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(99.31\%\)</EquationSource></InlineEquation> with an AUC of <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(99.97\%\)</EquationSource></InlineEquation> on the BSPCSI dataset, while showing improved performance compared with several baseline approaches and state-of-the-art deep learning models. To enhance interpretability and support clinical decision-making, explainable artificial intelligence techniques, including LIME and Grad-CAM, are integrated to highlight critical regions influencing predictions. Additionally, a web-based diagnostic tool is developed to enable real-time stroke prediction. The findings suggest that the proposed approach can serve as an effective and interpretable tool for automated stroke detection, with potential to enhance clinical diagnostic workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards trustworthy brain stroke diagnosis using a lightweight explainable deep learning framework for CT imaging

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月18日

查看全文
  • Md. Romzan Alom,
  • Muhammad Aminur Rahaman,
  • Md. Parvez Hossain,
  • Chayan Mondal,
  • Md. Nazmus Shakib,
  • Md. Ahsan Habib,
  • Md. Kamrul Hasan,
  • A. B. M. Shawkat Ali

摘要

Brain stroke occurs due to blockage or rupture in the cerebral blood supply and represents a critical medical emergency requiring rapid and accurate diagnosis. However, manual interpretation of CT scans is time-consuming and may delay clinical decision-making. To address this challenge, this study proposes the Deep Neural Brain Stroke Detection (DNBSD) system, a lightweight deep learning–based framework for automated stroke detection from CT images. The proposed model employs a task-specific convolutional neural network (CNN) architecture consisting of Conv2D, MaxPooling, Batch Normalization, Flatten and Dense layers, containing only 1.67 million trainable parameters and a computational complexity of 0.2973 GFLOPs, making it suitable for resource-constrained clinical environments. Additionally, preprocessing techniques including image resizing and normalization were applied to optimize performance. The model is trained and evaluated on two publicly available datasets: Brain Stroke CT Image Dataset (BSCI) and Brain Stroke Prediction CT Scan Image Dataset (BSPCSI), each divided into training, validation, and testing subsets. Experimental results demonstrate that the DNBSD system achieves high performance, with an accuracy of \(99.20\%\) and an AUC of \(99.96\%\) on the BSCI dataset, and an accuracy of \(99.31\%\) with an AUC of \(99.97\%\) on the BSPCSI dataset, while showing improved performance compared with several baseline approaches and state-of-the-art deep learning models. To enhance interpretability and support clinical decision-making, explainable artificial intelligence techniques, including LIME and Grad-CAM, are integrated to highlight critical regions influencing predictions. Additionally, a web-based diagnostic tool is developed to enable real-time stroke prediction. The findings suggest that the proposed approach can serve as an effective and interpretable tool for automated stroke detection, with potential to enhance clinical diagnostic workflows.