Click here to

Session: Adaptive Treatment Planning and Delivery [Return to Session]

Deep Learning Guided Segment-Maintained Adaptive Radiotherapy (SMART)

T Knewtson1*, G Szalkowski2, X Xu3, P Mavroidis4, J Dooley5, S Das6, J Lian7, (1) UNC Health Care, Chapel Hill, NC, (2) UNC Health Care, Chapel Hill, NC, (3) Rensselaer Polytechnic Institute, Troy, New York, (4) University of North Carolina, Chapel Hill, NC, (5) University of North Carolina, Chapel Hill, NC, (6) University of North Carolina, Chapel Hill, NC, (7) University of North Carolina, Chapel Hill, NC


SU-H300-IePD-F6-4 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Adaptive radiotherapy (ART) offers more optimal dosimetry and treatment outcomes for many patients with large anatomical changes than no action taken. However, the time required to generate a new patient plan and perform delivery quality assurance (DQA) hinders the wide use of ART in the conventional workflow. We propose a deep learning model and segment-shape maintained adaptive radiotherapy (SMART) approach for performing an efficient DQA-free ART.

Methods: Ten previously treated head and neck cancer patients were replanned for this study. Each patient was planned with 35 fractions to high-risk (70Gy), intermediate-risk (63Gy), and standard-risk (56Gy) PTV volume. Each plan used nine equally spaced 6MV IMRT fields. These patients experienced noticeable anatomical changes during the course of treatment. A new CT simulation was acquired and a new IMRT replan was created. A deep learning model was developed to predict the dose for the initial CT and replan CT. Three plans were created automatically with the guide of model prediction: plan on initial CT, plan on replan CT (with full optimization) and SMART plan on replan CT. The beams of the initial plan were directly copied onto the replan CT and dose was recalculated (DCT plan). This was used as a reference to simulate the delivered dose on treatment day with no intervention.

Results: Our model can predict the dosimetry of the deliverable plan accurately with no statistically significant differences of main dosimetric endpoints (p>0.05). The SMART plan significantly improves the target coverage on treatment anatomy without a clinically impactful dose increase for most organs-at-risk. Additionally, both replan methods can be completed within 10 minutes.

Conclusion: We developed a deep learning model to help create ART plans efficiently. The SMART approach improves the dosimetry of treatment day without the need for a new DQA.


Not Applicable / None Entered.


TH- External Beam- Photons: IMRT/VMAT dose optimization algorithms

Contact Email