Analysis of CNL-UNet for Efficient Biomedical Image Segmentation
Rifat Ahommed, Md. Badiuzzaman Shuvo, M. M. A. Hashem · 2021
Biomedical image segmentation has a lot of significance in disease detection and diagnosis. Many researchers proposed many deep learning architectures such as U-Net, SegNet, DualChexNet, DualANet for biomedical image segmentation. Recently, a novel lightweight architecture namely CNL-UNet is proposed for multimodal biomedical image segmentation. The CNL-UNet used the CNL module, res-path, and transfer learning for effective biomedical image segmentation. In this paper, we made an in-depth analysis of the CNL-UNet to prove its robustness on multimodal biomedical image segmentation. We experimented in detail with the various components of the CNL-UNet. We made eight combinations of the CNL-UNet by removing or adding these components and experimented with these eight combinations on the ultrasound and MRI datasets. The results show that the CNL module helps the model to reduce the false outputs and gain high precision, recall, and F1 score. The res-path has a remarkable contribution to provide precise segmentation and increasing the performance of the model. And also, transfer learning plays an important role in faster convergence and also in increasing the performance of the network. In essence, this experiment increases the confidence of the CNL-UNet regarding its remarkable performance on multimodal biomedical image segmentation.