Transfer Learning with ResNet-18 for Automated Lung Cancer Subtype Classification in Medical Imaging

This study proposes an automated classification framework based on the ResNet-18 convolutional neural network, which was pre-trained on the ImageNet dataset and subsequently fine-tuned using 15,000 lung tissue images from the LC25000 dataset. The dataset was stratified into training, validation, and testing sets—with proportions of 70%, 15%, and 15%, respectively—to maintain a balanced class distribution. Additionally, a systematic ablation study was conducted on four model configurations to evaluate the contributions of specific components: data augmentation, the use of a custom classifier head, and the implementation of differential learning rate scheduling. Experimental results demonstrate that the proposed full configuration achieved a test accuracy of 99.96% and a perfect macro-averaged F1-score of 1.0000. This performance was statistically superior to both the baseline model (McNemar's test: p = 0.023) and the configuration lacking differential learning rate scheduling (p = 0.041). Validation using Grad-CAM quantitative metrics revealed spatial coherence levels ranging from 96.04% to 100.00% across all classes, indicating that the model's regions of interest aligned with relevant histopathological characteristics. With 11.31 million parameters and an average inference time of only 2.80 ms per image, the proposed model not only delivers exceptional classification performance but also possesses sufficient computational efficiency for implementation in clinical environments with limited computational resources.