Lung cancer has been difficult to detect using CT scan due to class imbalance and high false negatives as well as for being difficult to interpret. This paper introduces DeepLungBalance, a holistic framework of evaluation of various deep learning architectures (VGG16, VGG19, ResNet50, MobileNet, and custom CNN) and solving these issues with advanced preprocessing techniques, dynamic balance of data, one class classification and class weighting techniques. Our integrated approach achieves an accuracy of 99.5%,F1=0.99, sensitivity =99% and specificity = 98.5% on MobileNetV2 and almost same result for VGG19. To enhance clinical interpretability, we combined Grad-CAM and LIME explanations of tumor-influencing areas in the CT scan in a form of visual heatmaps and pixel level explanations. Extensive evaluation in terms of cross-validation, showing progressive improvement from 72−86% on the imbalanced binary dataset to 99.5% using our integrated approach. A web-based system for deployment enables CT analysis in real-time with capacities to score, and explain the confidence of the analysis. These results demonstrate a step-wise combination of class imbalance adjustment and explainability resulting in a good, interpretable and potentially deployable solution for early detection of lung cancer.