Lung Lesion Classification Using VGG16-1D-LSTM and ResNet50-1D-LSTM: A Comparative Study of Basic vs Advanced Architectures

Lung cancer is a leading cause of cancer-related mortality worldwide, where early and accurate lesion classification is critical for improving patient outcomes. Traditional diagnostic methods relying on manual image interpretation are often subjective and time-consuming. To address these limitations, this study investigates the use of deep learning techniques for automated lung lesion classification. The objective is to compare the performance of two hybrid deep learning models: VGG16-1D-LSTM and ResNet50-1D-LSTM. A dataset of 250 samples, split for training (70%) and validation (30%) with Adam optimizer. Both hybrid models achieved 96% accuracy, demonstrating viability for clinical deployment. ResNet50-1D-LSTM offers superior precision-recall trade-offs, while VGG16-1D-LSTM excels in sensitivity. Their F1-scores were closely matched, 91.89% for VGG16-1D-LSTM and 92.31% for ResNet50-1D-LSTM. These results indicate a robust balanced performance from both models. The integration of the LSTM layer significantly improved the ResNet50-1D architecture's performance, increasing accuracy 84.67% to 96.00% and recall from 80.00% to 90.00%. While both models proved highly capable, the study noted limitations due to a small and less diverse dataset, which may impact external validity, and the significant computational resources required for model training. These findings demonstrate that integrating LSTM with well-established CNN architectures can substantially enhance lung lesion classification accuracy and reliability, potentially improving early diagnosis and patient outcomes in clinical practice.