Abstract
OBJECTIVES: Radiation-induced pneumonitis (RP) is a critical complication of radiotherapy in lung cancer patients, and its early detection remains a challenge due to the limited availability of annotated CT imaging data and the subtle nature of disease evolution. The objective of this study is to enhance the detection and localization of RP in CT by integrating advanced data augmentation, self-supervised learning, and synthetic data generation techniques. METHODS: A conditional Generative Adversarial Network (cGAN) was used to create synthetic RP images conditioned on lung segmentation masks to create anatomically plausible data for augmenting the training sets. The pipeline was created to possess double-stage self-supervised training with hierarchical pretext tasks to achieve robust features. RESULTS: The performance of the proposed framework, for a 5-fold cross-validation, has an average accuracy of 94.04%, precision of 92.06%, with a recall of 95.1%, an F1-score of 93.56%, and area under the curve of 95.94%. CONCLUSIONS: The model was demonstrated to possess superior performance and stability in RP detection and localization, which suggests potential clinical translation. ADVANCES IN KNOWLEDGE: The paper offers a novel fusion of cGAN-generated synthetic data, spatial attention, and contrastive learning to address RP detection in limited data. Interpretability is achieved by introducing Bayesian uncertainty estimation to provide translational value in clinical practice.