2202 - Harnessing Deep Learning on Fractional Dosimetry for Acute Genitourinary Toxicity Prediction Following Prostate SBRT

02:30pm - 04:00pm PT

Hall F

Screen: 3

POSTER

Presenter(s)

Xiangrong Qi, PhD - Dept. of Radiation Oncology, Los Angeles, CA

X. Qi¹, Y. Pang², J. Pham³, A. U. Kishan¹, L. Valle¹, M. L. Steinberg¹, M. C. Repka⁴, A. Wijetunga⁴, S. Sud⁴, and J. Lian⁴; ¹Department of Radiation Oncology, University of California, Los Angeles, CA, ²Department of Computer Science, University of North Carolina, Chapel Hill, NC, ³Department of Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, ⁴Department of Radiation Oncology, University of North Carolina, Chapel Hill, NC

Purpose/Objective(s): Stereotactic body radiotherapy (SBRT) is a widely used and highly effective treatment for localized prostate cancer. However, genitourinary (GU) toxicity remains a major concern, affecting patients' quality of life. Current predictive models primarily rely on planned dosimetry, with only a few studies incorporating treatment-delivered dosimetry. When delivery dosimetry is used, dose accumulation is typically performed via deformable registration, which introduces significant uncertainty due to registration inaccuracies and substantial changes in structure shape and volume. Moreover, adaptive therapy decisions for each fraction ideally depend on real-time predictive outcomes, yet this research direction remains largely unexplored. To address these challenges, we investigated the feasibility of early toxicity prediction by integrating planned and delivered dosimetry for each fraction directly. Our focus was on assessing whether data from the initial fractions could achieve satisfactory prediction accuracy and support adaptive treatment strategies.

Materials/Methods: We retrospectively analyzed 69 prostate cancer patients treated with MR-guided SBRT (40 Gy in five fractions) using a 0.35T MR-Linac. Dosimetric data, including dose-volume histograms (DVHs) from the original plan and delivered doses from each fraction, were compiled. Acute GU toxicity was classified using CTCAE criteria, with patients stratified into significant toxicity (Grade =2, 24.6%) and non-significant toxicity (75.4%). A unidirectional Long Short-Term Memory (LSTM) model and a Multi-Layer Perceptron (MLP) were employed to evaluate the associations between dosimetric features and GU toxicity in the two cohorts. Lasso regression was applied to identify key dosimetric endpoints and determine the optimal number of fractions required for accurate prediction. Predictive performance was assessed using the Area Under the Curve (AUC) metric and evaluated with a paired Student’s t-test.

Results: Incorporating dosimetric data from all five fractions improved prediction accuracy (AUC: 0.62 ± 0.04) compared to using planning dosimetry alone (AUC: 0.56 ± 0.03, p = 0.018). Notably, integrating the first two fractions with planning dosimetry significantly enhanced predictive performance (AUC: 0.74 ± 0.03, p < 0.001), whereas using fractions 4 and 5 instead reduced accuracy (AUC: 0.54 ± 0.04, p = 0.007). The most predictive dosimetric features were the urethra 41.5 Gy volume, urethra 41.2 Gy volume, and trigone 39.8 Gy volume.

Conclusion: This study demonstrates the improved prediction of acute GU toxicity by combining planned and delivered dosimetry for prostate SBRT treatment. Incorporating dosimetry from the initial fractions enhances prediction accuracy, enabling early treatment adjustments and potentially improving clinical outcomes.