Auto-redirect paused manually.
Efficient multi-class brain tumor classification from MRI images using InceptionResNetV2 bottleneck features and a lightweight dense classifier
Abstract
Abstract
The precise classification of brain tumors from magnetic resonance imaging (MRI) is essential for early diagnosis and effective treatment planning. This research introduces an efficient transfer learning framework that employs the InceptionResNetV2 architecture, pre-trained on ImageNet, as a frozen deep feature extractor, and couples it with a lightweight, strongly regularized fully connected classifier for four-class brain tumor classification (glioma, meningioma, pituitary tumor, and no tumor). The approach tackles the key challenges of limited annotated medical data and high computational cost by pre-computing 1,536-dimensional bottleneck features, so that only 158,954 parameters are trained. A benchmark dataset of 7,023 MRI images was employed for evaluation, partitioned into 70% training, 15% validation, and 15% test subsets. Evaluation measures including accuracy, precision, recall, F1-score, and the normalized confusion matrix were used to verify the model. Experimental findings indicate that the proposed model attains a test accuracy of 94.58%, a weighted precision of 0.9461, a weighted recall of 0.9458, and a weighted F1-score of 0.9459, with a validation accuracy of 95.18%, while requiring only about 12 minutes of training on a CPU-only workstation. The results demonstrate the robustness and practical significance of the proposed framework, establishing it as a computationally efficient alternative to fully fine-tuned deep models for automated brain tumor classification.
The precise classification of brain tumors from magnetic resonance imaging (MRI) is essential for early diagnosis and effective treatment planning. This research introduces an efficient transfer learning framework that employs the InceptionResNetV2 architecture, pre-trained on ImageNet, as a frozen deep feature extractor, and couples it with a lightweight, strongly regularized fully connected classifier for four-class brain tumor classification (glioma, meningioma, pituitary tumor, and no tumor). The approach tackles the key challenges of limited annotated medical data and high computational cost by pre-computing 1,536-dimensional bottleneck features, so that only 158,954 parameters are trained. A benchmark dataset of 7,023 MRI images was employed for evaluation, partitioned into 70% training, 15% validation, and 15% test subsets. Evaluation measures including accuracy, precision, recall, F1-score, and the normalized confusion matrix were used to verify the model. Experimental findings indicate that the proposed model attains a test accuracy of 94.58%, a weighted precision of 0.9461, a weighted recall of 0.9458, and a weighted F1-score of 0.9459, with a validation accuracy of 95.18%, while requiring only about 12 minutes of training on a CPU-only workstation. The results demonstrate the robustness and practical significance of the proposed framework, establishing it as a computationally efficient alternative to fully fine-tuned deep models for automated brain tumor classification.