Harnessing Convolutional Neural Networks for Diagnosing Post-SBRT Radiation-Induced Lung Injury From X-Ray CT

This study investigates the potential of Convolutional Neural Networks (CNNs) for diagnosing Radiation-Induced Lung Injury (RILI), a frequent complication of Stereotactic Body Radiation Therapy (SBRT), using post-therapy X-ray CT imaging. We retrospectively analyzed follow-up CT scans of 67 lung cancer patients from two institutions, employing a 3D CNN trained on image patches centered on the treatment isocenter. The CNN demonstrated effective discrimination between RILI and non-RILI cases, achieving an area under the ROC curve of 0.762 ± 0.029 and an F1 score of 0.789 ± 0.043 on the holdout test set. Its performance remained high in scans taken within three months post-SBRT and in scans of nodules under 2.5 cm diameter – both subsets of the data where diagnosis is challenging. These findings suggest that CNN features have good predictive power for diagnosing RILI post-SBRT, encouraging further research into integrating these features alongside other known predictors to enhance diagnostics.