Enhanced Pyramid Scene Parsing Network for Segmentation and Deep Dilated Residual CNN-based Sparrow Search Optimization for Efficient Liver Tumor Detection

Liver cancer, a highly fatal malignancy, is often diagnosed at a late stage, underscoring the need for early detection. While deep learning models address challenges like manual feature selection, they require substantial labeled data and lack transparency. DDRNet mitigates these issues through deep dilated convolutions, enabling precise feature extraction and reliable clinical application. The flow begins with input medical images, like computed tomography (CT) scans, which are subjected to preprocessing techniques to enhance their quality. Noise reduction and contrast enhancement are achieved using methods like the dynamic block size technique (DBST), local contrast modification (LCM), and contrast-limited adaptive histogram equalization (CLAHE), which increase image clarity and emphasize critical features, making tumor areas more detectable. The enhanced images are subsequently segmented using an improved pyramid scene parsing network (PSPNet), which correctly recognizes and delineates locations of interest, such as possible malignancies. A deep dilated residual convolutional neural network (DDRNet) is used in the classification step after segmentation to examine the segmented areas and identify complex spatial patterns. By adjusting DDRNet parameters, including the dropout rate, the sparrow search optimization algorithm (SSOA) is used to improve the system even more and guarantee precise and trustworthy categorization of tumor and non-tumor categories. These categorization findings are eventually produced by the system, which helps with the precise and early identification of liver malignancies like hepatocellular carcinoma (HCC). By automating critical diagnostic steps, this system significantly enhances diagnostic efficiency, reduces manual effort, and provides reliable support for medical professionals in liver tumor disease diagnosis. The proposed method obtains a Precision of 95.15%, false omission rate (FOR) of 4.85% and a Specificity of 95.24% for liver tumor disease detection, respectively. The proposed approach will enhance liver tumor detection by effectively capturing multi-scale spatial features and preserving fine details, resulting in more accurate diagnoses with minimal computational overhead.