Lung Cancer Classification from CT Images Using Ensemble Transfer Learning of VGG16, ResNet50V2 and MobileNetV2
Lung cancer remains one of the most common cancers worldwide, and how early it is caught shapes what treatment is possible and how long a patient lives. Computed tomography (CT) sits at the center of this process, and deep learning has become a standard tool for reading CT images, with strong results across medical image classification tasks. This paper proposes an ensemble approach that brings pre-trained VGG16, ResNet50V2, and MobileNetV2 together in a single framework, classifying lung CT images as benign, malignant, or normal. The dataset consists of 1946 CT images from 473 patients, collected at Atatürk University Hospital and labeled by a nuclear medicine physician. All three backbones were fully fine-tuned end-to-end, an approach that an ablation study confirmed outperformed both fully frozen and partially frozen configurations. Splits were made at the patient level throughout, so no patient appears in both training and validation. Five-fold cross-validation gave 89.42 ± 3.21% accuracy and a macro F1 of 0.865 ± 0.036; the same configuration, retrained on a separate patient-level split, reached 90.62% accuracy with a macro F1 of 0.874. The results indicate that ensemble transfer learning can serve as a supporting tool in lung cancer screening.