Transfer learning approaches for lung cancer detection using histopathological images: A comparative study of CNN models
Priyanka Rawat, Priyanka Kaushik · 2025
Lung cancer continues to be a major cause of cancer-related mortality globally, and early detection is the only way to survive. This research compares the performance of pretrained models such as DenseNet121, InceptionV3, VGG16, and VGG19 in classifying lung cancer from histopathological images. It compares the performance of CNN architectures VGG16, VGG19, Inception, and DenseNet121, classifying lung cancer into benign tissue, adenocarcinoma, and squamous cell carcinoma. DenseNet121 performed optimally with the highest accuracy at 95.64%, with a validation accuracy of 93.25% and lowest training and validation losses (0.18 and 0.21), showing improved generalization. Its dense connectivity connects each layer to all the following layers, enabling improved reuse of features and efficient passing of gradients in training. Inception also performed well with a validation accuracy of 91.98% and closely matched loss measures. VGG19 and VGG16 had validation accuracies of 89.03% and 87.76%, respectively, but showed some overfitting compared to DenseNet121 and Inception. Results indicate that contemporary architectures such as DenseNet121 and Inception perform better than traditional models (VGG16 and VGG19), showing their applicability in medical image classification. This research demonstrates the importance of advanced neural network architecture for efficient and accurate lung cancer detection.