Trident U-Net: An Encoder Fusion for Improved Biomedical Image Segmentation

Abstract

Image segmentation is a fundamental requirement in biomedical image analysis. Recent advances in deep learning have resulted in optimistic results on many biomedical image segmentation benchmark datasets. However, there is a need to propagate and improve biomedical image analysis due to significant variations in biomedical images in recent days. As with the rapid growth of technology in medical equipment, more variations of biomedical data of various tests/scans are being generated rapidly. There is a growing vacuum of quick technical analysis on these biomedical data. Biomedical image segmentation is at the forefront of them. Manual analysis results in outrageous efforts and costs because only biomedical experts can annotate effectively and are often subjected to human error. Various deep learning-based solutions are being created to leverage machine learning capabilities and effectively address this problem. In this paper, We propose a deep learning framework built on the U-Net architecture and use a combination (triplet) of encoder models to address the semantic segmentation of a biomedical image effectively. An extensive experiment of our proposed Trident U-Net architecture has been done using the CVC-612 and ETIS datasets. The overall performance of the proposed Trident U-Net on both datasets outperforms the existing reportedly best performing models. The results obtained from the ETIS dataset are 0.66, 0.68, and 0.59 for the evaluation metrics Recall, Dice Coefficient, and IoU Score, respectively. Similarly, the results achieved on the CVC-612 dataset are 0.91, 0.89, and 0.83 for the same evaluation metrics, respectively.