TransResUNet: Improving U-Net Architecture for Robust Lungs Segmentation in Chest X-rays

Medical image segmentation is regarded as an important component in a computer-aided diagnosis (CAD) system as it directly affects overall system performance. In this paper, we propose a new fully convolutional encoder-decoder model for lung segmentation named TransResUNet. We developed this architecture improving the state-of-the-art U-Net model. As part of the improvement to the classical U-Net, we introduced a pre-trained encoder, a special skip connection and a post-processing module in the proposed architecture. As a result, the proposed model outperformed the baseline U-Net model by 97.6% vs 94.9% considering the dice coefficient and 98.5% vs. 96.8% in terms of accuracy. The proposed TransResUNet achieved this feat with about 24% fewer parameters than the baseline U-Net. The implementation (based on Keras) of our proposed model is publicly available at https://sakibreza.github.io/TransResUNet with additional resources.