An Optimized EfficientNet-B3 Based Lung Cancer Detection Framework with Stratified Splitting and Deployment-Ready Torch Script Integration

Lung cancer still remains one of the major causes of cancer deaths worldwide, which in turn indicates the need for efficient, accurate, and automated diagnostic support systems. In this context, this paper proposes a completely automated system for the diagnosis of lung cancer using chest computed tomography scans, specifically using a deep learning approach and the EfficientNet-B3 model as the backbone for the proposed system. To enhance the generalization of the proposed system, despite the limited dataset of 1,097 images, stratified data splitting, optimized data augmentation, and adaptive learning rate schedules are employed in the proposed system. To ensure class balance in the dataset, a stratified data splitting approach of 70-20-10 was used for the proposed system, where the dataset was divided into 70% for training, 20% for validation, and 10% for testing. During the training process, the system converged well, achieving a high accuracy of 99.48%, a minimal loss of 0.0197, and a maximum validation accuracy of 99.09%. Additionally, the generalization performance of the proposed system, in terms of accuracy and loss, on the test dataset was found to be high, achieving a maximum accuracy of 98% and a minimum loss of 0.0566, thereby validating the effectiveness of the proposed system in accurately classifying lung cancer, benign, and normal images. Thus, the proposed system, specifically using the EfficientNet-B3 model, optimized preprocessing, and a well-designed system, presents a promising and efficient solution for the automated diagnosis of lung cancer.