BRAIN TUMOR SEGMENTATION USING U-NET-BASED DEEP LEARNING MODELS

When examining segmentation methods developed for brain tumor detection, a significant gap exists in the comparative performance analysis of standard and hybrid U-Net models. This study systematically evaluates four Convolutional Neural Network (CNN) models to address this gap, particularly regarding the practical effectiveness of integrating complex, pre-trained encoders: Classic U-Net, U-Net+VGG-16, U-Net+ResNet50, and U-Net++. The BraTS 2019 dataset was used for training and testing the models (with 5-fold cross-validation), and the Figshare dataset was used for additional validation, performing tumor segmentation and generating predictions. U-Net++ outperformed both the conventional U-Net and the hybrid models, indicating its higher segmentation performance under the conditions evaluated in this study. While U-Net++ achieved superior results in Dice coefficient, sensitivity, specificity, and Jaccard Index values (0.920, 0.890, 0.973, and 0.850, respectively), the U-Net+VGG-16 hybrid model (0.697, 0.659, 0.862, and 0.535, respectively) and the U-Net+ResNet50 hybrid model (0.585, 0.550, 0.800, and 0.414) produced lower Dice scores. When we examine the primary reasons for this poor performance output, we identify and discuss these as the incompatibility of 2D encoder blocks with 3D medical data and the lack of unoptimized transfer learning. We believe this finding is of critical importance for future model designs. Our study demonstrates that the U-Net++ architecture, with its dense and intertwined skip connections, provides a more effective solution than simple hybrid models built with powerful classifiers such as VGG-16 and ResNet50. This allows us to obtain extremely important results for precise brain tumor segmentation.