Investigating the Impact of Data Fusion on the ResUNet Model for Liver Tumor Segmentation

Abstract

Medical image segmentation is employed to separate regions in images based on color and shape differences for disease diagnosis or localization of pathological areas. It can be performed manually or automatically. Automatic segmentation methods leverage machine learning and deep learning techniques, with domain-specific models developed to enhance performance; U-Net-based architectures can achieve effective results even with limited and imbalanced medical datasets. However, U-Net models may face training challenges such as vanishing gradients in deep networks, which can be addressed by ResNet architectures providing deeper structures. The hybrid ResUNet combines the segmentation capabilities of U-Net with the residual connections of ResNet, thus exploiting the advantages of deep networks while mitigating training difficulties. In this study, for automatic liver tumor segmentation, the hybrid ResUNet was applied to channel-based fused data obtained using Principal Component Analysis (PCA) and Discrete Wavelet Transform (DWT). Channel-based data fusion preserves the unique and distinctive patterns of each channel, enriching feature representations, and both PCA-and DWT-based fusion methods transform the data into different spaces, enhancing the model's ability to differentiate various structures. The results demonstrate that both methods achieve comparable performance, yielding similar dice similarity coefficient values across two different datasets.