2068 - Development of a Markerless Real-Time Monitoring Method for Radiation Therapy in Prostate Cancer

Purpose/Objective(s): Fiducial markers are commonly used for real-time localization in prostate cancer radiation therapy; however, their implantation poses risks such as bleeding and infection. This study aims to develop and evaluate a non-invasive, high-precision markerless localization method by utilizing naturally occurring prostatic calcifications and deep learning techniques.

Materials/Methods: We analyzed cone-beam CT (CBCT) images acquired for image-guided radiation therapy (IGRT) from 25 prostate cancer patients treated at our institution between 2015 and 2018. For each case, 9 to 24 CBCT scans were obtained, with five randomly selected for testing and the remaining (4 to 19) used for training. The target for localization was the largest calcification within the planning target volume (PTV). To augment the training data, the target calcification was artificially shifted by up to 1 cm, generating 50 CBCT images per case. Digital Reconstructed Radiographs (DRRs) were generated from CBCT images and used as input data, with the ground truth defined as the centroid of the calcification. A deep learning model, Swin UNETR, was developed to predict the probability map of calcification locations. The model was trained to process five consecutive DRRs (each at 1-degree increments) as a five-channel input, predicting the centroid position of each calcification. During validation, a "statistical approach" was implemented to exclude outliers: if three or more out of five predictions clustered within 2 mm, their mean position was used as the final output; otherwise, the output was discarded.

Results: The proposed method achieved a mean absolute deviation (MAD) of 0.98 ± 0.21 mm across all cases. In 18 cases with a calcification volume of =0.05 cc (High group), MAD was 0.51 ± 0.07 mm, achieving submillimeter localization accuracy. In contrast, in seven cases with a calcification volume <0.05 cc (Low group), MAD increased significantly to 2.17 ± 0.53 mm (p = 0.0049). Angular accuracy analysis was performed by dividing DRR projections into six groups at 60° intervals. In the High group, no significant differences in accuracy were observed across angles (Friedman test, p = 0.84). However, in the Low group, localization accuracy decreased particularly in lateral directions (60°–119° and 240°–299°), indicating a significant difference in MAD across angles (p = 0.019).

Conclusion: This study developed a novel method for high-precision localization of calcifications in kV images for prostate cancer radiation therapy. Localization accuracy decreased in lateral directions when the calcification volume was <0.05 cc. However, for calcification volumes =0.05 cc, submillimeter precision was achieved regardless of projection angle. This approach represents a promising alternative for markerless position monitoring in prostate cancer radiation therapy.