Hypopharyngeal Cancer Segmentation Using Convolution Neural Network: A Comparative Analysis

Image segmentation methods based on deep learning can help doctors segment the region of interest (ROI) in medical images rapidly. Hypopharyngeal cancer (HPC) is a rare cancer, so there are fewer Magnetic Resonance Images (MRIs) for HPC. MRIs of HPC often have problems of uneven brightness. Therefore, it is a major challenge to use deep learning to build a semantic segmentation network for HPC. To solve the problem, we choose ResNetl8 (pretrained on ImageNet dataset) as the encoder and compare the effects of different convolution neural networks as decoders. Compare with other methods, i.e., U-Net, DeepLabV3+, U-Net++, and EfficientUNet++, ResUNet++ achieves splendid prediction performance and outperforms other methods on HPC datasets, which achieves the highest Dice score of 77.34%.