Background Immune checkpoint inhibitors (ICIs) are a cornerstone of modern oncological treatments, particularly in the management of various cancers through immunotherapy. Despite their clinical success, ICIs are often associated with several immune-related adverse events (irAEs), among which pneumonitis is particularly significant due to its potential severity. Accurately differentiating ICI-induced pneumonitis from normal pulmonary findings and other interstitial lung abnormalities (ILAs) at an early stage remains a critical challenge in patient management. Early detection and precise differentiation are essential for timely and appropriate therapeutic interventions, which can significantly alter patient outcomes.

Methods We analyzed baseline CT scans from 349 patients receiving ICI therapy to develop a deep learning predictive model. Utilizing an autoencoder-based method with a 3D Res-Unet50 architecture, we enhanced detection capabilities for early pneumonitis by applying data augmentation techniques like brightness normalization and pixel shuffling. The model was trained on normal cases using anomaly detection via mean squared error (MSE) distribution and then tested on pneumonitis cases. Initial training involved a 3D residual attention U-Net optimized with L1 loss and stochastic gradient descent, focusing on anomaly detection through image reconstruction and MSE analysis to evaluate performance.

Results The proposed deep learning model significantly outperformed traditional models in identifying high-risk pneumonitis patients with ICI, achieving 84.54% accuracy, 82.86% precision, 85.29% sensitivity, 83.78% specificity, 84.06% F1-score, and an 84.96% AUC. In contrast, the clinical model scored substantially lower in all metrics, with 64.79% accuracy, 62.86% precision, and a 60.65% AUC. Radiomics and filter-based radiomics models showed improvement over the clinical model, with the latter achieving a precision of 84.38% and specificity of 86.49%. However, these models still fell short of the deep learning model’s performance, highlighting its superior predictive capabilities compared to traditional, clinically based assessments.

Conclusions The predictive model we developed uses deep learning and autoencoder-based feature extraction to effectively distinguish between normal cases and those at risk of developing ICI-induced pneumonitis. Integrating this model into clinical workflows could enhance decision-making, improve patient monitoring, and potentially reduce the incidence and severity of pneumonitis. This advancement in imaging analytics could improve patient outcomes and underscores the value of machine learning in managing irAEs.