PatchResNet: Multiple Patch Division–Based Deep Feature Fusion Framework for Brain Tumor Classification Using MRI Images
Taha Muezzinoglu,Nursena Baygin,Ilknur Tuncer,Prabal Datta Barua,Mehmet Baygin,Sengul Dogan,Turker Tuncer,Elizabeth Emma Palmer,Kang Hao Cheong,U. Rajendra Acharya
DOI: https://doi.org/10.1007/s10278-023-00789-x
IF: 4.903
2023-02-19
Journal of Digital Imaging
Abstract:Modern computer vision algorithms are based on convolutional neural networks (CNNs), and both end-to-end learning and transfer learning modes have been used with CNN for image classification. Thus, automated brain tumor classification models have been proposed by deploying CNNs to help medical professionals. Our primary objective is to increase the classification performance using CNN. Therefore, a patch-based deep feature engineering model has been proposed in this work. Nowadays, patch division techniques have been used to attain high classification performance, and variable-sized patches have achieved good results. In this work, we have used three types of patches of different sizes (32 × 32, 56 × 56, 112 × 112). Six feature vectors have been obtained using these patches and two layers of the pretrained ResNet50 (global average pooling and fully connected layers). In the feature selection phase, three selectors—neighborhood component analysis (NCA), Chi2, and ReliefF—have been used, and 18 final feature vectors have been obtained. By deploying k nearest neighbors (kNN), 18 results have been calculated. Iterative hard majority voting (IHMV) has been applied to compute the general classification accuracy of this framework. This model uses different patches, feature extractors (two layers of the ResNet50 have been utilized as feature extractors), and selectors, making this a framework that we have named PatchResNet. A public brain image dataset containing four classes (glioblastoma multiforme (GBM), meningioma, pituitary tumor, healthy) has been used to develop the proposed PatchResNet model. Our proposed PatchResNet attained 98.10% classification accuracy using the public brain tumor image dataset. The developed PatchResNet model obtained high classification accuracy and has the advantage of being a self-organized framework. Therefore, the proposed method can choose the best result validation prediction vectors and achieve high image classification performance.
radiology, nuclear medicine & medical imaging