CNN-Driven Deep Transfer Learning for Lung Cancer Classification (CNN-DTL-LCC) Using ResNet Architectures: An Early Diagnosis Robust Framework

Abstract

Lung cancer is the most common cause of cancer death in developing countries like India. The lung cancer patient survival rate will be increased with early diagnosis, but lack of effective diagnostic techniques for early detection and most of the failure treatment for metastatic disease make it difficult. The conventional diagnostic methods remain affected by subjectivity, radiological variability, and intensive computational demands in clinical workflows. This research proposed a Convolutional Neural Network-Driven Deep Transfer Learning for Lung Cancer Classification (CNN-DTL-LCC), which is a reliable multi-class classification model implemented through three CNNs: “ResNet-50, ResNet-101, and EfficientNet-B0”. The dataset utilized in this work contains four categories of lung tissue images: Adenocarcinoma (ADC), Large Cell Carcinoma (LCC), Normal Lung Tissue (NLT), and Squamous/Epidermoid Cell Carcinoma (SCC/ECC). In order to enhance model generalization, this four-class histopathology dataset was preprocessed using Random rotation, Horizontal/Vertical Flipping, Random Scaling and Brightness Jitter. The entire pre-trained model was fine-tuned with the Adam optimizer (learning rate of 1×10⁻⁴, batch size of 16, for 10 epochs), by exchanging the real fully connected layer with a custom four-class head. The proposed CNN-DTC-LCC model achieves the maximum performance with ResNet-50, attaining an accuracy of 96.83%, an F1-score of 96.2%, and an AUC of 0.981. Grad-CAM visualizations validate that the model precisely annotates disease-specific sections, improving interpretability and trust in clinical diagnosis. Analysis of the confusion matrix validates a stronger discriminatory ability compared to the other classes, especially within malignant subtypes. In the context of the CNN-DTL-LCC framework, the current results position ResNet-50 as an efficient, scalable, and clinically feasible solution for early lung cancer diagnosis that offers a solid foundation for improving suitable arrangements for AI-driven diagnostic issues, particularly within low-resource healthcare contexts.