Pre-Processing of Computed Tomography for Liver Tumor Segmentation using Multiscale Residual Learnable Shuffle Attention V-Net

In liver tumor segmentation, pre-processing plays crucial role in preparing Computed Tomography (CT) scans for accurate performance. Segmentation involves delineating tumor regions from CT for supporting clinical diagnosis by identifying tumor boundaries, size, and shapes. However, accurate liver tumor segmentation remains primary challenge due to low contrast, irregular shapes, and overlapping boundaries with surrounding tissue. Without effective pre-processing to minimize noise, enhance contrast, and preserve structural details, segmentation struggle to differentiate subtle tumor that leads to suboptimal performance. In this research, Multiscale Residual Learnable Shuffle Attention V-Net (MRLSAV-Net) with pre-processing method is proposed for liver tumor segmentation using CT scans. In traditional V-Net, MRLSA is incorporated by capturing multiscale contextual features and emphasizing most appropriate channel and spatial information. LSA refines feature representation by focusing on tumor regions which leads to more accurate performance. Adaptive Weighted Median Filter (AWMF) is used to minimize noise by preserving tumor boundaries whereas Contrast Limited Adaptive Histogram Equalization (CLAHE) enhance local contrast making subtle liver lesions more distinguishable. Hence, MRLSAV-Net with pre-processing achieves high Dice score of 0.9794 and 0.9553 on LiTS17 and 3DIRCADb datasets compared to existing methods like Multi-Attention Network (MANet). AWMF+CLAHE obtains high PSNR values of 41.56 dB and 44.25 dB on both datasets.