3590 - Association Between SBRT Dose Distribution Profiles and Disease Recurrence in Early-stage NSCLC Patients: Implementation of AI-based Recurrence Prediction Tool

Purpose/Objective(s): Stereotactic Body Radiation Therapy (SBRT) is a standard curative treatment option for patients with medically inoperable non-small cell lung cancer (NSCLC). Despite technological advancements in SBRT delivery, the risk of disease recurrence remains clinically significant. Early identification of recurrence risk is therefore essential, as it may support timely clinical decision-making and optimization of subsequent treatment strategies. This study aims to develop an artificial intelligence (AI)-based model for predicting recurrence using SBRT dose distribution maps.

Materials/Methods: This study included 287 patients with early-stage NSCLC treated with SBRT, with prescribed doses ranging from 30-60 Gy delivered in 1-5 fractions. The cohort comprised 51% male patients, with a median age of 73 years (range: 41-96 years). The primary endpoint was disease recurrence (n = 63), including local, distant, and combined failures. To develop the AI-based prediction model, patient-specific dose distribution maps were analyzed. Quantitative imaging features were extracted from dose profiles within a contour defined by 50% of the maximum prescribed dose. To ensure methodological consistency, the CT slice exhibiting the highest cumulative pixel intensity was selected for each patient. Feature extraction was performed using a deep architecture consisting of five residual blocks followed by a dense neural network with five hidden layers. Subsequently, a prediction network comprising seven fully connected hidden layers (with 50, 100, 250, 150, 100, 50, and 30 neurons, respectively) was implemented to predict recurrence.

Results: Model performance was assessed using 10-fold cross-validation to ensure robustness and reduce overfitting. The proposed framework achieved a mean AUC of 0.82 (±0.05), demonstrating strong discriminative ability. Sensitivity and specificity were 0.81 (±0.13) and 0.83 (±0.16), respectively, indicating balanced detection of both recurrence and non-recurrence cases. The weighted average F1-score was 0.83 (±0.09), and the weighted precision was 0.88 (±0.04), reflecting reliable classification performance across outcome classes. Additional evaluation metrics confirmed model stability: balanced accuracy reached 0.82 (±0.06), and the geometric mean of sensitivity and specificity was 0.81 (±0.06). Results indicate consistent and well-balanced predictive capability.

Conclusion: This study presents a deep neural network framework for predicting recurrence in early-stage NSCLC patients treated with SBRT. Dose map-derived imaging features showed strong predictive value, underscoring the role of spatial dose characteristics in recurrence risk assessment. This approach may aid early risk stratification, personalized surveillance, and post-SBRT treatment planning. Further external validation is needed to confirm generalizability and clinical utility.