MSCAUNet-3D: Multiscale Spatial Channel Attention 3D-UNet for Lung Carcinoma Segmentation on CT Image

Abstract

The lung segmentation process plays a vital role in diagnosing lung carcinoma. Segmentation techniques segment the lung region and remove the borders, blood vessels, and void spaces in the CT images. For segmentation, segmenting the highlighted vital features and suppressing the unwanted features is important. In the paper, we proposed the new segmentation techniques combined with an attention mechanism to achieve accurate segmentation. In this model, we introduced the multiscale spatial and channel attention mechanism with the 3D-UNet model named MSCAUNet-3D. this model performs two stages: pre-processing and segmentation. In pre-processing, adaptive Histogram Equalization (AHE) and Gaussian Adaptive Bilateral Filter (GABF) are utilized for removing noise and enhancing the image. In segmentation, we introduce the MSCAUNet-3D for accurate segmentation. To evaluate this model, Dice Coefficient (DC), Jaccard Similarity Coefficient or Index (JI), and Relative Absolute Volume Difference (RAVD) performance measures are utilized. The proposed model yields 91.4, 90.4, and 89.4 in DC, JI, and RAVD, respectively, which shows that the proposed model outperforms the other models.