Lung and Colorectal Malignancies rank high on the list of causative factors for cancer-associated death across the globe, mainly because of their late diagnosis. Untimely detection of lung cancer itself is a consequence of symptoms emerging when it is already too advanced for any form of intervention. Colorectal malignancies exhibit comparable problems of early detection, often culminating in dismal survival rates when diagnosed too late. These aforementioned problems lead to increased demand for accurate, reliable, and timely detection systems to allow for better treatment and prognosis. The present study offered a novel approach toward cancer prediction, integrating preprocessing, attention mechanisms, and ensemble learning techniques. The data preprocessing pipeline ensures clean, balanced datasets, thus increasing model reliability. An attention mechanism attends to important features in the imaging data, thus improving feature extraction and representation. The proposed ensemble model integrates DenseNet and ResNet architectures so that their complementary strengths can be utilized to capture both hierarchical and fine-grained information. This method achieved an amazing 99.74% accuracy with great promise in cancer prediction. This emphasizes the benefits of using preprocessing, attention-based feature extraction, and ensemble deep learning models in increasing cancer detection accuracy.