Enhancing Liver Tumor Segmentation Performance with Sequential and Parallel Hybridization of Pretrained Models

Medical imaging plays a role, in identifying and treating liver tumors in patients with liver diseases. It is vital for tumor detection. Planning surgical procedures. Understanding the details of medical images is crucial in this process. U-net,a neural network known for its encoder decoder architecture is widely used for segmentation purposes due to its ability to precisely locate objects, in medical images. Transfer learning proves its value in situations where there is a shortage of labeled medical imaging data. This scenario can pose a challenge, to the training of networks. This study delves into utilizing trained models such, as ResNet50, ResNet101 and ResNet152 in a U framework for liver tumor segmentation. These pre trained models were originally trained on datasets like ImageNet. Are now being used to extract essential features for the task at hand. While VGG16 has an architecture that can capture spatial information effectively the ResNet models, with deep residual connections excel at capturing hierarchical features. Using these models as encoders has greatly enhanced the segmentation performance of U-NET. Additionally the study introduces a technique that blends feature maps from trained models to improve segmentation accuracy significantly. The superior performance results, from merging the strengths of models using both parallel blending methods. These hybrid models surpass the effectiveness of trained networks showing a notable enhancement, in key metrics. The results indicate that incorporating trained models, into the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$U$</tex> net framework enhances liver tumor segmentation by providing a wider range of features leading to a more precise and accurate delineation of tumors. This approach presents a direction for enhancing medical imaging segmentation tasks and delivering efficient tools, for clinical decision making.