Augmented self attention network for liver tumour segmentation of CT images

Computed Tomography (CT) scans are frequently employed for the deep learning based liver tumour segmentation for the accurate diagnosis of liver conditions. An Augmented Self-Attention (ASA) network can effectively capture long-range information, addressing the primary limitation of Residual-Unet methods, which focus on local neighbourhoods and consequently miss global context. Here ASAneXt50 is proposed that makes use of map of convolutional features with attention feature maps in ResNeXt, leading to the development of an adept method for liver tumour segmentation. Experimental findings reveal that the proposed model achieved better liver and tumour segmentation results in Three Dimensional Image Reconstruction for Comparison of Algorithm Database (3DIRCADb) and Liver Tumour Segmentation (LiTS) dataset. The architecture achieved a higher level of Dice score (DSC) – 0.923, Intersection-Over-Union (IoU) – 0.913, Volumetric Overlap Error (VOE) – 0.086 and Relative Volume Difference (RVD) – 3.85 for liver tumour segmentation which is remarkable.