Liver tumor kills millions of people annually, and standard diagnostic methods include liver CT scans. However, noisy and pointless features in huge datasets might reduce the efficacy of deep learning (DL) systems. The feature selection, for the pre-processing technique, has the potential to decrease database and enhance the classification of accuracy by the selection of the most significant feature. Furthermore, this study presents a robust evaluation of the methodology which significantly reduces the cost associated with evaluation while enhancing the effectiveness of the feature selection method. This approach aims to overcome the limitation of the expensive solution which has been analyzed in the wrapper strategy. To overcome feature selection challenges, this study developed a Hybrid Spotted Hyena Optimization (HS-WO) with Walrus Optimization (WaOA), which successfully identified the ideal subset with the greatest number of pertinent qualities. The liver CT-Scan datasets are utilized for acquiring biomedical liver tumor data, while bicubic-interpolation technique has been used for the noise reduction. We employed the DCGAN generative modelling technique for facilitating the data augmentation process. This selected liver characteristics has been assessed by utilizing a CNN-LSTM that discovered both normal and aberrant features in biological liver tumor data. The suggested classifier achieved 99.8% sensitivity, 99.3% specificity, 99.14% F1-score, and 99.6% accuracy in the liver image database.