Deep learning-based prediction of interfractional anatomic variations in prostate cancer radiotherapy.
Authors
Affiliations (3)
Affiliations (3)
- Department of Medical Physics and Biomedical Engineering, School of Medicine, Shahid Beheshti University of Medical Sciences, Tehran, Iran.
- Department of Medical Physics and Biomedical Engineering, School of Medicine, Shahid Beheshti University of Medical Sciences, Tehran, Iran. [email protected].
- Radiation Biology Research Center, Iran University of Medical Sciences, Tehran, Iran. [email protected].
Abstract
Interfractional prostate displacement challenges the accuracy of the delivered dose during radiotherapy treatment. Variation in bladder volume is a key driver of this movement. A deep learning (DL) model was developed to predict daily megavoltage computed tomography (MVCT) based on bladder volume and treatment fraction, enabling patient-specific anatomic estimation in prostate radiotherapy. This retrospective study analyzed 700 MVCT scans from prostate cancer patients treated with tomotherapy. The bladder was manually contoured on all MVCT scans, and its volume was calculated. A customized three-dimensional (3D) U‑Net model was trained to generate synthetic MVCT images using the kilovoltage computed tomography (KVCT), bladder volume, and fraction number as inputs. The model's performance was evaluated with data from 84 held-out MVCTs, using image similarity metrics including structural similarity index (SSIM), normalized cross-correlation (NCC), Dice similarity coefficient (Dice), mean absolute error (MAE), and mean squared error (MSE). Additionally, anatomic accuracy was assessed for bladder and prostate contours, applying the Dice similarity and MSD. The model demonstrated accurate MVCT predictions, evidenced by a mean SSIM ranging from 0.76 to 0.80, NCC from 0.84 to 0.89, a Dice of 0.97, and MAE and MSE values between 0.05 and 0.06 and 0.010 and 0.014, respectively. The predicted anatomy enabled bladder contouring with a mean Dice of 0.83 and mean surface distance (MSD) of 1.64 mm and prostate contouring with a mean Dice of 0.92 and MSD of 0.48 mm. Changes in bladder volume showed moderate correlation with bladder Dice, with a correlation coefficient (r = -0.50), and with MSD (r = 0.64), while the prostate contour metrics exhibited weak correlations. The proposed deep learning (DL) framework demonstrated promising capabilities for anatomic prediction and contour generation using noninvasive input features. Within the context of helical radiotherapy delivered on a Radixact X9 system (Accuray Inc., Madison, WI, USA), where daily MVCT acquisition is standard practice, this approach provides a patient-specific method for estimating the anatomic configuration on the day of treatment based on the planning KVCT and bladder volume information.