Evaluation of Deep Learning Architectures for Pulmonary CT Lesion Classification Highlighting Diagnostic Performance

Early identification of pulmonary lesions is a critical factor in enhancing patient prognosis and survival rates. This study systematically evaluates the diagnostic performance of five deep learning architectures, ConvNeXt Base, ResNet50, EfficientNetV2 Small, InceptionV4, and Xception, for the three-class categorization of Computed Tomography (CT) scans into Benign, Malignant, and Normal categories. Utilizing the public IQ OTH NCCD dataset, we applied a transfer learning approach with ImageNet weights, complemented by a robust training pipeline incorporating dynamic data augmentation and early stopping to mitigate overfitting and ensure model generalization. Model efficacy was rigorously assessed on an independent test set using accuracy, precision, recall, and F1-score metrics. Experimental results indicate that InceptionV4 emerged as the most reliable architecture, achieving an overall accuracy of 0.988 and a macro-averaged F1-score of 0.976. Notably, this model demonstrated perfect sensitivity for the pathologically critical malignant class, achieving a recall rate of 1.00, thereby prioritizing clinical safety. These findings confirm that advanced neural networks can serve as dependable secondary opinion systems for clinicians. Given its superior sensitivity and balanced diagnostic profile, InceptionV4 represents a promising candidate for integration into automated lung cancer screening workflows to improve diagnostic precision.