Evaluating CNN Architectures and Hyperparameter Tuning for Enhanced Lung Cancer Detection Using Transfer Learning

Abstract

Accurate lung cancer detection is vital for timely diagnosis and treatment. This study evaluates the performance of six convolutional neural network (CNN) architectures, ResNet-50, VGG-16, ResNet-101, VGG-19, DenseNet-201, and EfficientNet-B4, using the LIDC-IDRI dataset. Models were assessed both in their base forms and with transfer learning. The dataset consisted of 460 × 460 × 3 pixel images categorized into squamous cell carcinoma (SCC), normal benign, large cell carcinoma (LCC), and adenocarcinoma (ADC). Performance metrics were computed, including accuracy (99.47% for the custom CNN), precision (99.50%), recall (98.37%), AUC (99.98%), and F1-score (98.98%) during training. However, overfitting was observed in the validation phases. Transfer learning models showed better generalization, with DenseNet-201 achieving a top validation accuracy of 96.88% and EfficientNet-B4 of 96.53%. Hyperparameter tuning improved the models’ generalization capabilities, maintaining high accuracy while reducing overfitting. This study highlights the effectiveness of transfer learning, particularly DenseNet-201, in enhancing automated lung cancer detection systems. Future work will focus on expanding datasets and exploring additional augmentation techniques to further refine model performance in clinical settings.