Automated Brain Tumor Identification in Biomedical Radiology Images: A Multi-Model Ensemble Deep Learning Approach

Abstract

Brain tumors (BT) represent a severe and potentially life-threatening cancer. Failing to promptly diagnose these tumors can significantly shorten a person’s life. Therefore, early and accurate detection of brain tumors is essential, allowing for appropriate treatment and improving the chances of a patient’s survival. Due to the different characteristics and data limitations of brain tumors is challenging problems to classify the three different types of brain tumors. A convolutional neural networks (CNNs) learning algorithm integrated with data augmentation techniques was used to improve the model performance. CNNs have been extensively utilized in identifying brain tumors through the analysis of Magnetic Resonance Imaging (MRI) images The primary aim of this research is to propose a novel method that achieves exceptionally high accuracy in classifying the three distinct types of brain tumors. This paper proposed a novel Stack Ensemble Transfer Learning model called “SETL_BMRI”, which can recognize brain tumors in MRI images with elevated accuracy. The SETL_BMRI model incorporates two pre-trained models, AlexNet and VGG19, to improve its ability to generalize. Stacking combined outputs from these models significantly improved the accuracy of brain tumor detection as compared to individual models. The model’s effectiveness is evaluated using a public brain MRI dataset available on Kaggle, containing images of three types of brain tumors (meningioma, glioma, and pituitary). The experimental findings showcase the robustness of the SETL_BMRI model, achieving an overall classification accuracy of 98.70%. Additionally, it delivers an average precision, recall, and F1-score of 98.75%, 98.6%, and 98.75%, respectively. The evaluation metric values of the proposed solution indicate that it effectively contributed to previous research in terms of achieving high detection accuracy.

Enhancing brain tumor detection in MRI with a rotation invariant Vision Transformer

“…However, a pertinent limitation is the model's focus on binary classification, whereas other studies in the field often tackle multiclass scenarios, presenting a more nuanced challenge (Mahmud et al, 2023 ; Natha et al, 2024 ). Additionally, the incorporation of rotational patch embedding introduces an extra layer of complexity, but this does not translate to an increase in hyperparameters due to the rotational operations.…”

“…Further advancements in the field include Natha et al (2024) multi-model ensemble deep learning approach for automated brain tumor identification, and (Gade et al, 2024) optimized Lite Swin transformer model combined with a barnacle mating optimizer for hyper-parameter tuning, achieving higher classification results and processing efficiency compared to existing transfer learning methods. Liu et al (2023) employed an ensemble of ViTs for glioblastoma tumor segmentation, exemplifying the power of combining multiple ViT models to improve segmentation outcomes.…”