Semantic Segmentation of Brain Tumor MRI Images Using ResUNet, ResNet-50, and DenseNet121: A Comparative Study

Brain tumor segmentation from Magnetic Resonance Imaging (MRI) is an important task in tumor analysis, treatment planning, and disease monitoring. However, manual delineation of tumor regions is time-consuming, subjective, and prone to inter-observer variability. This study presents a comparative evaluation of three deep learning-based 2D semantic segmentation models: ResUNet, ResNet-50, and DenseNet121, using the BraTS 2020 multimodal MRI dataset. The MRI images were preprocessed using rotation, histogram equalization, Gaussian blur, and intensity normalization. The three models were evaluated under a standardized experimental framework using identical training, validation, and test partitions. Segmentation performance was assessed for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) using Accuracy, Dice coefficient, Intersection over Union (IoU), HD95, Precision, Sensitivity, and Specificity across three independent runs. In the slice-level evaluation, ResUNet achieved the highest mean Dice for WT (0.7936 ± 0.0573), TC (0.7309 ± 0.0984), and ET (0.6589 ± 0.0774), together with the highest mean IoU and precision across all three tumor regions. ResUNet achieved the lowest HD95 for WT, whereas ResNet-50 obtained the lowest HD95 for TC and ET and the highest sensitivity for WT. ResUNet achieved the highest sensitivity for TC and ET and the highest specificity across all three tumor regions. DenseNet121 demonstrated a comparatively lower segmentation performance, particularly for ET. Additionally, a separate patient-level evaluation was conducted using 100 MRI subjects who were not included in the original 20-patient dataset. Volumetric segmentation performance was assessed by combining 2D slice-wise predictions into complete MRI volumes. ResUNet achieved the highest patient-level Dice for WT (0.6569 ± 0.0204), TC (0.4598 ± 0.0236), and ET (0.4676 ± 0.0268). Furthermore, slice-level tumor areas were calculated in pixels from predicted and ground-truth segmentation masks to assess tumor area estimation. Overall, ResUNet demonstrated the strongest and most consistent segmentation performance across slice-level and patient-level evaluations.