<p>Magnetic Resonance Imaging (MRI) is a widely used medical imaging modality due to its non-invasive and radiation-free nature. It is especially effective for visualizing soft tissues such as the brain. However, current MRI systems often struggle to acquire high-resolution images due to limitations like patient movement, short acquisition time, and signal-to-noise ratio challenges. These limitations reduce diagnostic accuracy, particularly in conditions such as stroke. The aim of this study is to enhance MRI image resolution and improve stroke disease classification accuracy by developing an artificial intelligence (AI)-based deep learning (DL) framework, thereby supporting more effective clinical diagnosis and treatment planning. Primarily, low-resolution MRI images are obtained from an Acute Ischemic Stroke MRI dataset, which comprises approximately 1002 stroke MRI images in.png and. jpg formats, along with 1008 corresponding control files. These images passed through the proposed Graph Convolution Residual-Generative Adversarial Network (GCR-GAN) to generate high-resolution outputs. The GCR-GAN modifies traditional Generative Adversarial Network (GAN) architecture by incorporating a graph-based residual learning block for enhanced super-resolution performance. Following this, stroke lesion segmentation is carried out on the high-resolution MRI images using Deep Object Segmentation Network (Deep O-SegNet), a hybrid model developed by integrating DeepJoint Segmentation with an Object Segmentation Network (O-SegNet). To increase dataset diversity and robustness, image augmentation techniques, such as flipping, shearing, shifting, and random rotation are applied. Feature extraction is then performed using Dual Tree- Fuzzy Local Binary Pattern (DT-FLBP) combined with entropy, and Gray-Level Co-Occurrence Matrix (GLCM) features. Subsequently, stroke classification is executed using the HX-ShuffleNet classifier, which is fine-tuned using Chronological Gold Rush Optimization (CGRO). HX-ShuffleNet is a deep learning model created by fusing ShuffleNet and Xception architectures, while CGRO is a novel optimization algorithm developed by integrating the Gold Rush Optimizer (GRO) with a chronological search concept. The performance of the proposed method is analyzed using the Acute Ischemic Stroke MRI dataset comprises a total of 2010 MRI images, including 1002 abnormal and 1008 normal cases. For experimentation, the dataset is divided into 70% training (1407 images), 15% testing (301 images), and 15% validation (302 images) subsets. The proposed AI framework demonstrated significant performance in super-resolution and stroke classification tasks. The GCR-GAN achieved optimal super-resolution performance, attaining a Peak Signal-to-Noise Ratio (PSNR) of 49.58 dB, a Second derivative like measure of enhancement (SDME) of 50.48 dB, and a Structural Similarity Index Measure (SSIM) of 0.92. Similarly, the CGRO-HX-ShuffleNet model achieved a True Positive Rate (TPR) of 96.49%, overall accuracy of 96.18%, True Negative Rate (TNR) of 95.42%, precision of 95.778, and F1-score of 96.418, outperforming existing approaches. The integration of AI and machine learning (ML) methods, specifically the GCR-GAN for super-resolution and CGRO-HX-ShuffleNet for classification, greatly enhances diagnostic precision in stroke detection from MRI images. This approach addresses resolution and diagnostic limitations in current MRI systems and can be extended to other clinical imaging applications.</p>

A Novel GCR-GAN for MRI Image Super-Resolution in Stroke Disease Classification

期刊论文
发表日期：

2026年3月5日

查看全文

S. E. Viswapriya,
D. Rajeswari

摘要

Magnetic Resonance Imaging (MRI) is a widely used medical imaging modality due to its non-invasive and radiation-free nature. It is especially effective for visualizing soft tissues such as the brain. However, current MRI systems often struggle to acquire high-resolution images due to limitations like patient movement, short acquisition time, and signal-to-noise ratio challenges. These limitations reduce diagnostic accuracy, particularly in conditions such as stroke. The aim of this study is to enhance MRI image resolution and improve stroke disease classification accuracy by developing an artificial intelligence (AI)-based deep learning (DL) framework, thereby supporting more effective clinical diagnosis and treatment planning. Primarily, low-resolution MRI images are obtained from an Acute Ischemic Stroke MRI dataset, which comprises approximately 1002 stroke MRI images in.png and. jpg formats, along with 1008 corresponding control files. These images passed through the proposed Graph Convolution Residual-Generative Adversarial Network (GCR-GAN) to generate high-resolution outputs. The GCR-GAN modifies traditional Generative Adversarial Network (GAN) architecture by incorporating a graph-based residual learning block for enhanced super-resolution performance. Following this, stroke lesion segmentation is carried out on the high-resolution MRI images using Deep Object Segmentation Network (Deep O-SegNet), a hybrid model developed by integrating DeepJoint Segmentation with an Object Segmentation Network (O-SegNet). To increase dataset diversity and robustness, image augmentation techniques, such as flipping, shearing, shifting, and random rotation are applied. Feature extraction is then performed using Dual Tree- Fuzzy Local Binary Pattern (DT-FLBP) combined with entropy, and Gray-Level Co-Occurrence Matrix (GLCM) features. Subsequently, stroke classification is executed using the HX-ShuffleNet classifier, which is fine-tuned using Chronological Gold Rush Optimization (CGRO). HX-ShuffleNet is a deep learning model created by fusing ShuffleNet and Xception architectures, while CGRO is a novel optimization algorithm developed by integrating the Gold Rush Optimizer (GRO) with a chronological search concept. The performance of the proposed method is analyzed using the Acute Ischemic Stroke MRI dataset comprises a total of 2010 MRI images, including 1002 abnormal and 1008 normal cases. For experimentation, the dataset is divided into 70% training (1407 images), 15% testing (301 images), and 15% validation (302 images) subsets. The proposed AI framework demonstrated significant performance in super-resolution and stroke classification tasks. The GCR-GAN achieved optimal super-resolution performance, attaining a Peak Signal-to-Noise Ratio (PSNR) of 49.58 dB, a Second derivative like measure of enhancement (SDME) of 50.48 dB, and a Structural Similarity Index Measure (SSIM) of 0.92. Similarly, the CGRO-HX-ShuffleNet model achieved a True Positive Rate (TPR) of 96.49%, overall accuracy of 96.18%, True Negative Rate (TNR) of 95.42%, precision of 95.778, and F1-score of 96.418, outperforming existing approaches. The integration of AI and machine learning (ML) methods, specifically the GCR-GAN for super-resolution and CGRO-HX-ShuffleNet for classification, greatly enhances diagnostic precision in stroke detection from MRI images. This approach addresses resolution and diagnostic limitations in current MRI systems and can be extended to other clinical imaging applications.