The multi-stage classification of lung cancer based on CT images is challenging due to morphological similarities between subtypes, variations in image contrast, and class distribution imbalances in medical datasets. Although deep transfer learning approaches have shown promising results in CT-based lung cancer detection, most previous studies focused on binary classification and applied static contrast enhancement, thus limiting adaptive feature learning. This study presents a contrast-adaptive and class-balanced transfer learning framework using EfficientNetB1 for multi-stage classification of lung cancer. Gamma correction is integrated directly into the augmentation pipeline as an on-the-fly preprocessing function, so contrast variations can be dynamically studied during the training process. The proposed approach also applies an increase in input resolution to 240×240 pixels, as well as a calibrated partial fine-tuning strategy on the last 50 layers of EfficientNetB1 to improve domain adaptation while maintaining computational stability. To reduce bias due to class imbalance, class weighting is used in the optimization process. The model was evaluated on a CT dataset with four classes, Adenocarcinoma Ib, Large Cell Carcinoma IIIa, Squamous Cell Carcinoma IIIa, and Normal, achieving an overall accuracy of 97% with a macro F1-score of 0.97, indicating balanced performance across classes. The proposed framework enhances the ability to discriminate between subtypes that have visual similarities while maintaining strong generalization capabilities. These findings suggest that the integration of contrast-based adaptive augmentation with calibrated learning transfer and class-balancing-based optimization can improve the resilience and stability of multi-stage classification of CT-based lung cancer.