Skip to content
Conference Open access

A Transfer-Learning Approach for Multi-Class Brain Tumor Classification Using a Novel Parameter-Optimized Alex Net-GRU Model

Sep 2026 · 2026 7th International Conference on Smart Electronics and Communication (ICOSEC) · pp. 999-1004 · 0 citations · 18 references

Abstract

Brain tumors require timely and reliable identification because their location and type can strongly influence clinical management. Owing to its ability to capture detailed anatomical structures without exposing patients to ionizing radiation, magnetic resonance imaging (MRI) has become a commonly preferred technique for examining the brain anatomical information without ionizing radiation. However, manual interpretation of a large number of MRI images can be laborious and may be affected by observer variability. This study investigates a lightweight deep-learning framework that combines a pretrained AlexNet feature extractor with a gated recurrent unit (GRU) classifier for four-class brain tumor recognition. Four types of diagnoses are covered in this experiment, namely, glioma, meningioma, pituitary tumor, and non-tumorous images. The experiments were conducted based on Version 2 of the Brain Tumor MRI Dataset which is found on Kaggle platform, consisting of 7,200 MRI images. Out of the total number of images, 5,600 images were used in developing and training the models, and the other 1,600 images were left for evaluation purposes. The images were resized to have a spatial resolution of 224 * 224 before being normalized using the ImageNet pre-training values of the AlexNet model. The proposed network uses AlexNet to learn spatial representations and a GRU-based feature-processing stage before the final four-class decision. feature-refinement stage prior to producing the final four-class output. The model was trained with the use of the Adam optimizer having the learning rate of 0.0001, the batch size of 32, and the dropout rate of 0.1 for 20 epochs. During the training, the model was able to reach an accuracy of 98.40% on the validation set. Moreover, it was able to attain precision, recall, and F1-score of 98.50%. A custom VGG16 implementation was used as a comparison model and obtained 95.50% accuracy, 95.25% precision, 95.00% recall, and 95.00% F1-score. These findings suggest that the proposed hybrid architecture strikes a favorable trade-off between predictive performance and the computational resources required to achieve it. LIME was additionally used to provide an interpretable view of the regions influencing model predictions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.