An Improved VGG Based Deep Learning Framework for Glioma Classification in Brain MRI Images

Introduction: Early detection of brain tumors can play a significant role in treatment planning. In brain magnetic resonance imaging (MRI), there are considerable similarities between healthy and cancerous tissues, making accurate tumor detection a major challenge. Furthermore, manual classification of these images is time‑consuming and prone to human error; therefore, developing automated classification methods can provide substantial support to physicians.

Methods: In this study, a novel deep neural network architecture is proposed for automatic classification of brain tumors in MRI images. In the proposed method, the VGG‑16 deep neural network architecture has been redesigned. To preserve the spatial resolution of features in deeper layers and to enable detection of small tumor regions as well as identification of ambiguous boundaries, the max‑pooling layer in the original VGG‑16 architecture was removed.

Results: The proposed method was evaluated using the BraTS2020 dataset with a 10‑fold cross‑validation approach. The evaluation results, based on the metrics Accuracy, Precision, Recall, Specificity, F1 Score, and AUC, showed that the proposed architecture achieved better performance compared to the original VGG‑16 neural network. Specifically, it attained values 0.9512 ± 0.0035, 0.9739 ± 0.0026, 0.9524 ± 0.0044, 0.9490 ± 0.0051, 0.9630 ± 0.0027, and 0.9514 ± 0.0044, respectively, for the mentioned metrics.

Conclusion: The results obtained from the proposed method in this study confirm that preserving spatial information and identifying ambiguous boundaries can play a significant role in improving classification accuracy. In addition, the applied modifications led to faster convergence and reduced over‑downsampling.