International Journal of Advances in Electrical Engineering
2023, Vol. 4, Issue 1, Part A
Enhancing brain tumor classification accuracy with advanced deep learning on MRI data
Author(s): Anubhuti Singh and Raghavendra Patidar
Abstract: Building a trustworthy and precise deep learning-based system for MRI-based brain tumor classification and comparing its performance to other deep learning methods is the goal of this effort. The goal is to classify brain tumors into three categories: glioma, meningioma and pituitary using publicly available datasets. Examining binary categorization (tumor vs. no tumor) is also part of the plan. We extracted the MRI pictures from the Figshare dataset, which contains data that is both multi-class and binary-labeled. There were three sets of data: training (70%), validation (15%) and testing (15%). The data had already undergone preprocessing steps such as normalization, scaling and real-time data augmentation. The four models were DenseNet201, ResNetV2, InceptionV3 and SVM (the baseline). After training with ImageNet weights via transfer learning, deep learning models were fine-tuned and classified using softmax. To train the models, we utilized the Adam optimizer in conjunction with the categorical cross-entropy loss function. To prevent them from becoming overfit, we coupled early stopping with learning rate scheduling. To evaluate performance, we calculated the F1-score, Accuracy, Precision and Recall using the test set. Test accuracy was 99.97%, precision was 96.78% and recall was 97.78% for DenseNet201, making it the top model out of the ones we looked at. Though its test accuracy was slightly lower at 94.78%, InceptionV3's recall was quite high at 98.08%. In contrast, ResNetV2 achieved a well-rounded performance with a recall of 95.07% and an accuracy of 93.45%. When compared to the results, the baseline SVM model's accuracy of 85.74% was much lower. According to confusion matrices and segmentation overlays, DenseNet201 learned the most complex patterns for precise tumor localization and classification. Based on the results, deep learning architectures—and DenseNet201 in particular—are much superior to more traditional approaches, such as support vector machines (SVM), when it comes to MRI-based brain tumor classification. These findings demonstrate the potential for improved clinical neuro-oncology diagnosis through the use of transfer learning and CNN-based models.
DOI: 10.22271/27084574.2023.v4.i1a.88
Pages: 100-106 | Views: 67 | Downloads: 20
Download Full Article: Click Here
How to cite this article:
Anubhuti Singh, Raghavendra Patidar. Enhancing brain tumor classification accuracy with advanced deep learning on MRI data. Int J Adv Electr Eng 2023;4(1):100-106. DOI: 10.22271/27084574.2023.v4.i1a.88
