Scholarlink

The detection and classification of brain tumours is one of the most urgent problems in medical diagnostics because of the varying complexity of brain structures, intensity irregularities, and noise in MRI images. Radiologists are challenged to manually analyse vast numbers of MRI scans in real time in a clinical setting, which results in misdiagnosis, delayed treatment, and lower survival rates. Conventional image processing and single-purpose deep learning models, such as CNNs and standard DNNs, are unable to achieve trade-offs among precision, computational efficiency, and generalisation, particularly in the presence of diverse tumour morphologies and limited-sized annotation datasets. To overcome these shortcomings, this paper proposes an optimised Hybrid Deep Neural Network-VGG16 (DNN-VGG16) architecture combined with Swin-UNet segmentation to analyse brain tumours efficiently and automatically. In the preprocessing phase, a Linear Smoothing Filter is used to remove noise and improve clarity. Swin-UNet is then used to perform segmentation to ensure that the tumour regions are well delineated, and thereafter, a comprehensive feature extraction procedure, which considers statistical features, PCA-NGIST, Local Gabor XOR Pattern (LGXP), Speeded-Up Robust Features (SURF) and CNN-based deep features, is done to ensure that the tumour characteristics are well represented. The obtained features are then presented to the proposed Hybrid DNN-VGG16 classifier, which leverages VGG16’s representational capability and DNN’s flexibility to classify tumour and non-tumour areas with high accuracy. The BRATS 2018 and Figshare databases have been experimentally verified to achieve improved accuracy of 89.7, precision of 87.9, recall of 88.5, and F1 score of 88.2, compared to the existing CNN and EfficientNet models. These findings demonstrate that the proposed hybrid model has a significant positive effect on diagnostic reliability, reduces computational load, and provides an efficient, real-time scaling architecture for medical decision-making in brain tumour detection and classification.

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

Hybrid DNN-VGG16 Framework for Brain Tumor Segmentation and Classification Using MRI Imaging: an Optimized Deep Learning Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ch. Dhanunjaya Rao,
  • A. Kumaravel

摘要

The detection and classification of brain tumours is one of the most urgent problems in medical diagnostics because of the varying complexity of brain structures, intensity irregularities, and noise in MRI images. Radiologists are challenged to manually analyse vast numbers of MRI scans in real time in a clinical setting, which results in misdiagnosis, delayed treatment, and lower survival rates. Conventional image processing and single-purpose deep learning models, such as CNNs and standard DNNs, are unable to achieve trade-offs among precision, computational efficiency, and generalisation, particularly in the presence of diverse tumour morphologies and limited-sized annotation datasets. To overcome these shortcomings, this paper proposes an optimised Hybrid Deep Neural Network-VGG16 (DNN-VGG16) architecture combined with Swin-UNet segmentation to analyse brain tumours efficiently and automatically. In the preprocessing phase, a Linear Smoothing Filter is used to remove noise and improve clarity. Swin-UNet is then used to perform segmentation to ensure that the tumour regions are well delineated, and thereafter, a comprehensive feature extraction procedure, which considers statistical features, PCA-NGIST, Local Gabor XOR Pattern (LGXP), Speeded-Up Robust Features (SURF) and CNN-based deep features, is done to ensure that the tumour characteristics are well represented. The obtained features are then presented to the proposed Hybrid DNN-VGG16 classifier, which leverages VGG16’s representational capability and DNN’s flexibility to classify tumour and non-tumour areas with high accuracy. The BRATS 2018 and Figshare databases have been experimentally verified to achieve improved accuracy of 89.7, precision of 87.9, recall of 88.5, and F1 score of 88.2, compared to the existing CNN and EfficientNet models. These findings demonstrate that the proposed hybrid model has a significant positive effect on diagnostic reliability, reduces computational load, and provides an efficient, real-time scaling architecture for medical decision-making in brain tumour detection and classification.