Due to differences in the intensity of the Computed Tomography (CT) scan image, none of the several lung cancer detection methods are enough to detect accurate malignancy. The NSCLC-Radiomics database is used to compile the dataset, which includes 300 individuals. U-Net, a deep learning model, is used for segmentation in the validation process. To achieve better results when segmenting CT scans, pre-processing procedures are used to boost the original image's poor contrast. Furthermore, the skip connections are enhanced with a selected kernel unit to acquire multi-scale features with variable receptive field sizes attained by means of soft attention. Using the Modified War Search Optimizer (MWSO) to fine-tune the parameters of the proposed modified U-Net, the classification accuracy is improved. Lastly, a Convolutional Neural Network (CNN) ensemble pre-trained classifier is used to detect the classes of liver tumors. Classification of liver tumors is accomplished using the suggested DL models, which comprise Mobile Net, VGG16, Xception, EfficientNetB7, and ResNet50. Dice similarity and sensitivity are two measures that measure the intersection of ground facts and forecasts; they are used to evaluate the segmentation method's performance. Obtaining an accuracy of 97% is a testament to the efficacy of the proposed method in correctly identifying and segmenting lung tumors.