A Comparative Evaluation of Different Pretrained Models for Automatic Segmentation of Gliomas

The automatic segmentation of gliomas is a laborious task due to the complexity of the gliomas' morphology, size and location. Most of the available deep learning-based techniques exploit architectures from scratch which requires extensive GPU memory and other computational resources. Moreover, these architectures fail to identify extremely minute core and enhancing regions of High Grade Glioma (HGGs). So, this work presents the utilization of the four reliable pre-trained models from the families of MobileNet and EfficientNet for the effective segmentation of three glioma sub-constituents. Both these families present the lightweight architectures which not only minimizes the trainable parameters but also elevates the segmentation efficiencies. The experimental analysis is performed on Multimodal brain tumor segmentation challenge (BraTS 2020) database and it is indicated that EfficientNet B4 produces highest values of dice coefficient as 0.968, 0.971 and 0.944 for the complete, core and enhancing regions of gliomas respectively.