Abstract
Lung cancer remains a leading global health issue. This research used artificial intelligence to enhance lung nodule characterisation within CT screening. Using transfer learning techniques, the study emphasised the concordance between AI predictions and human-assigned labels. Leveraging a pre-trained VGG19 model, two independent datasets of lung nodule images were investigated. Overall, high agreement and classification performance were seen in the training dataset. However, a drop in model reliability on unseen data was noted, highlighting the challenges of perfect agreement between AI and human interpretations. Still, high sensitivity, TPR, and AUC values affirmed the efficacy of the paradigm chosen. Despite areas of potential improvement, the approach demonstrated the promising role AI and transfer learning could play in early lung cancer detection. Future improvements in AI interpretability may yield even more robust, clinically viable prediction models.
