3468 - Optimizing Deep Learning Models: CNN-RNN Augmentation with Grad-CAM Analysis for Predicting Pneumonitis from Pre-Treatment CT Images

Z. Zou¹, P. T. Teo², A. Yalamanchili³, and M. Abazeed³; ¹Northwestern University, Chicago, IL, ²Department of Radiation Oncology, Northwestern University, Feinberg School of Medicine, Chicago, IL, ³Department of Radiation Oncology, Northwestern University Feinberg School of Medicine, Chicago, IL

Purpose/Objective(s): Radiation therapy is crucial for lung cancer treatment. However, incident irradiation of surrounding healthy tissues could lead to side effects such as radiation-induced pneumonitis—inflammation that can cause permanent lung damage. Determining a patient’s susceptibility to pneumonitis prior to treatment is not currently possible and is a critical unmet clinical need. We hypothesize that deep-learning (DL) models using pre-treatment CT images can identify patients at risk of radiation-induced pneumonitis. Materials/

Methods: We analyzed 1,168 IRB-approved CT images from two medical institutions, comprising 64 cases of grade =2 pneumonitis. Strategic sampling was employed to balance the outcome representation. The whole lung area was selected as our Region of Interest (ROI) and each slice was segmented to obtain the mask for the ROI. The segmented mask and corresponding original image were input into our custom DL model. The model used a two-stage classification: (1) a two-layer convolutional neural network (CNN) extracts features from each CT slice and (2) a bidirectional recurrent neural network (RNN) combines these features with adjacent slice data to predict the pneumonitis risk for the patient. The dataset was divided into a ratio of 80:20 for training and cross-validation. Grad-CAM was used to highlight regions predictive of pneumonitis.
Results: Our custom two-layer CNN achieved a slice-level ROC-AUC of 0.6713. This performance improved significantly upon aggregating slice-level predictions for patient-level diagnostics, with the AUC score increasing to 0.75 through static aggregation methods. The application of the RNN for dynamic aggregation further elevated the AUC to an impressive 0.8285. The ensemble CNN-RNN architecture effectively captures local features and temporal dependencies between CT slices, enabling accurate prediction of radiation-induced pneumonitis. The models robust performance across multi-institutional data underscores its generalizability and potential for widespread clinical application in personalizing lung cancer radiotherapy treatment plans.
Conclusion: We developed a parsimonious DL model using an ensemble CNN-RNN architecture that can accurately predict radiation pneumonitis using pre-treatment CT images alone. These results signify a promising step towards personalized treatment strategies and improved outcomes for patients undergoing radiotherapy to the lung. This study has implications for clinical practice, potentially enabling physicians to identify high-risk patients and modify treatment plans to mitigate the risk of radiation-induced pneumonitis.