Automated U-Net-ConvMixer Attention method for Lung Segmentation

Abstract

Accurate lung segmentation in chest X-rays is vital for diagnosing various pulmonary pathologies. While U-Net architectures and their derivatives have achieved success in medical applications, their local convolution operations inherently limit their ability to capture global contextual information. In this work, we present a novel ConvMixer-based model for lung segmentation. Inspired by the ConvMixer architecture, this model effectively extracts both local and global features from lung images. To improve our segmentation results, we proposed a post-processing step in order to eliminate weakly contributing features from the segmentation. We evaluate our model on two publicly available chest X-ray datasets, Shenzhen and Montgomery, demonstrating superior performance compared to state-of-the-art segmentation methods. Notably, our final model achieves an accuracy of 97.52% and an IoU of 92.71%. These results suggest the proposed ConvMixer-based model as a promising approach for lung segmentation with the potential to contribute to improved diagnosis of various lung diseases.