The proper delineation of the multi-modal MRI of the gliomas and of the brain is important in diagnosis, surgical planning, and monitoring of treatment in neuro-oncology. There is no doubt that convolutional neural networks CNNs) have become the standard of medical image segmentation. But still there is a limited amount of comparative evidence on how different architectures model tumor morphology. This paper is a systematic evaluation of eight CNN-based segmentation models namely U-Net, ResU-Net, Attention U-Net, RAU-Net, Feature Pyramid Network (FPN), SegNet, BiSeNet and DeepLabV3+ using the BraTS 2020 dataset. Each model was trained with the same preprocessing and hyperparameter settings, without augmentation or pre-training, to determine architecture-based effects. Performance was evaluated on a variety of accuracy measures (mean Intersection over Union, precision, recall, specificity, AUC) and efficiency measures (inference time). Consistently, FPN and ResU-Net outperformed ground truth (mIoU ≈0.896), strong recall (>0.92), and sub-milliseconds inference time, indicating better ability to detect infiltrative tumor margins and still be clinically deployable. Attention U-Net presented the best accuracy (0.96), minimizing false positives, but lightweight BiSeNet did not capture tumor boundaries despite being able to run fast. This work contributes insights into the encoding of glioma morphology by deep learning models, and identifies directions to pursue biologically-informed model development, by connecting the architectural performance of the deep learning model to the clinically-relevant error profiles.