Retrospective Clinical Evaluation of a Decision-Support Software for Adaptive Radiotherapy of Head and Neck Cancer Patients

Sébastien Gros, Anand P. Santhanam, Alec M. Block, Bahman Emami, Bryan Lee, Cara Joyce

TL;DR

Researchers evaluated RTapp, a new decision-support platform that automates the monitoring of anatomical changes in head and neck cancer (HNC) patients during radiation. The software successfully identified that 67% of patients required treatment plan adaptations and demonstrated the ability to predict dose-limit violations up to four days in advance.

What Problem Does This Paper Address?

In radiotherapy for head and neck cancer, a patient’s anatomy changes significantly over the standard 6–7 week treatment course due to weight loss, tumor shrinkage, or shifts in positioning. Adaptive Radiotherapy (ART)—the process of adjusting the radiation plan to account for these changes—is the gold standard for ensuring the tumor receives the full dose while sparing healthy organs like the parotid (salivary) glands.

However, traditional "offline" ART is incredibly labor-intensive. It requires physicians and physicists to manually review daily images, re-contour anatomy, and recalculate doses—a process that can take days. Because of this "resource bottleneck," many clinics only replan when anatomical changes are visually obvious, which may be too late. This paper evaluates whether an automated software (RTapp) can provide real-time, quantitative data to trigger these adaptations proactively.

Methods

The authors performed a retrospective study on 27 HNC patients using the RTapp (SegAna) platform. The software follows a three-step automated workflow:

  1. Deformable Image Registration (DIR): It automatically maps the original planning CT and structures onto daily Cone-Beam CT (CBCT) images using an optical flow-based algorithm.
  2. Dose Deformation: It "warps" the planned dose onto the daily anatomy to estimate the actual dose delivered that day.
  3. Trend Analysis & Prediction: It calculates cumulative doses and uses a linear predictive model to forecast dosimetric parameters (like PTV coverage or parotid mean dose) for the next four fractions.

The study used specific "Warning" and "Adaptation" thresholds (e.g., a 2% drop in PTV coverage or a 10% increase in organ-at-risk dose) to determine when the software would have triggered a clinical intervention.

Datasets

The study utilized a retrospective cohort of 27 HNC patients treated with VMAT (Volumetric Modulated Arc Therapy) at Loyola University Medical Center in 2019.

  • Initial Pool: 81 patients.
  • Selection Criteria: Patients required daily kV-CBCT imaging and complete PTV/parotid volume coverage within the imaging field of view.
  • Diversity: The cohort included various HNC sites, including the pharynx, oral cavity, and larynx.

Key Insights and Ideas

  • The "Two-Phase" Violation Pattern: The study identified two distinct reasons for dose violations. Early violations (within the first few fractions) were typically caused by sub-optimal immobilization, such as head tilts or shoulder misalignment. Late violations (after the second quarter of treatment) were driven by gradual biological changes like weight loss.
  • Real-Time Feasibility: RTapp processed a single fraction (93 CBCT images and ~30 structures) in less than one minute. This suggests that quantitative ART monitoring can be moved from a "days-long" offline task to a "near-instant" check at the treatment console.
  • Proactive vs. Reactive Planning: By predicting violations four fractions ahead, the software gives the clinical team a "buffer" to create a new treatment plan before the current plan becomes dosimetrically unsafe.
  • The Shoulder Problem: A significant finding was that inconsistent shoulder positioning (often missed in standard visual checks) significantly impacted the dose to inferior cervical lymph nodes, highlighting the need for better immobilization or more sensitive monitoring.

Key Results

  • High Need for Adaptation: 18 out of 27 patients (67%) violated at least one dose constraint by the end of treatment.
  • Target Coverage: 12 PTVs showed a mean coverage decrease of 6.8% ± 2.9%. Six of these were flagged for adaptation at a median of fraction 6.
  • Organ Sparing: 17 parotid glands exceeded mean dose constraints, with a median increase of +2.60 Gy.
  • Prediction Accuracy: The linear model was reasonably accurate, with mean differences between predicted and actual values of -2.7% for PTV coverage and 0.3 Gy for parotid dose. Accuracy improved as the patient approached the "threshold-crossing" fraction.

Limitations and Future Work

  • Small Sample Size: The study was limited to 27 patients, which led to wide confidence intervals in the prediction model's accuracy.
  • Field of View (FOV) Constraints: Current CBCT systems have a limited FOV. If part of the tumor or an organ is "cut off" in the daily image, the software cannot accurately calculate the dose for that structure.
  • Linear Model Simplicity: The current prediction relies on linear regression. The authors suggest that Machine Learning (ML) models incorporating more variables (like weight loss velocity or baseline anatomy) could provide more robust forecasts.
  • Lack of Independent Recalculation: The version of the software used could not export deformed CTs, meaning the authors couldn't compare RTapp’s dose estimates against a secondary "gold standard" dose engine.

The authors conclude that while prospective trials are needed, RTapp provides a viable path toward making adaptive radiotherapy a standard, automated part of daily clinical practice.

Summary updated

Abstract

Integrated in an ART clinical workflow, RTapp aids in predicting whether specific treatment would require adaptation up to four fractions ahead of time.
BibTeX
@inproceedings{gros2022retrospective,
  author = {Sébastien Gros and Anand P. Santhanam and Alec M. Block and Bahman Emami and Bryan Lee and Cara Joyce},
  title = {{Retrospective Clinical Evaluation of a Decision-Support Software for Adaptive Radiotherapy of Head and Neck Cancer Patients}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.777793},
}
Figure