Enhanced MRI brain tumor segmentation with DNet and hybrid dice-weighted cross-entropy loss

Localization of brain tumors via magnetic resonance imaging (MRI) is critically important for diagnosis, treatment, and surgical operations. This study proposes DNet, an optimized and highly efficient deep learning framework for the automatic detection and segmentation of brain tumors from MR images. On the basis of the classical UNet structure, DNet reduces the risk of overfitting by balancing the integration of batch normalization and dropout layers; it also achieves higher accuracy in small tumor regions by combining Dice loss and weighted cross-entropy (WCE) loss. To improve the model's robustness and predictive stability, an ensemble and test-time augmentation (TTA) were applied during testing. The dataset used in the study consisted of 1664 MR images containing three tumor classes: glioma, meningioma, and pituitary. Data preprocessing and augmentation steps, including contrast correction, reflection, and cropping, were applied to the images. The model was trained via the five-fold cross-validation method; the mean intersection-over-union (mIoU) and accuracy were used for evaluation. The experimental results demonstrate improved performance over the classical UNet. The DNet model achieved competitive performance compared with existing approaches in the literature, reaching 99.6% accuracy and a test mIoU value of 0.8617 on a three-class dataset. Furthermore, ablation studies revealed that reducing the number of classes increased the model's mIoU by approximately 3%. In conclusion, the proposed DNet framework provides a robust and efficient framework for medical image segmentation. In addition to the primary dataset, consistent performance was also observed on an independent dataset, supporting the robustness of the proposed approach. Nevertheless, further validation on diverse and multi-center datasets is required for real-world clinical deployment.