Initial Evaluation of a Novel Cone-Beam CT-Based Semi-Automated Online Adaptive Radiotherapy System for Head and Neck Cancer Treatment - A Timing and Automation Quality Study

Abstract

Introduction A novel on-line adaptive radiotherapy (ART) system based on O-ring linear accelerator (LINAC) and cone-beam CT (CBCT) was evaluated for treatment and management of head & neck (H&N) cancer in an emulated environment accessed via remote desktop connection. In this on-line ART system, organs-at-risk (OARs) and target contours and radiotherapy (RT) plans are semi-automatically generated based on the patient CBCT, expediting a typically hours-long RT planning session to under half an hour. In this paper, we describe our initial experiences with the system and explore optimization strategies to expedite the process further. Methods We retroactively studied five patients with head and neck cancers, treated 16-35 fractions to 50-70 Gys. For each patient, on-line ART was simulated with one planning CT and three daily CBCT images taken beginning, middle, and end of treatment (tx). Key OAR (mandible, parotids, and spinal cord) and target (planning target volume (PTV) = clinical target volume (CTV) + 3 mm margin) contours were auto-generated and adjusted as needed by therapist/dosimetrist and attending physician, respectively. Duration of OAR contouring, target contouring, and plan review was recorded. Key OAR auto-contours were qualitatively rated from 1 (unacceptable) - 5 (perfect OAR delineation), and then quantitatively compared to human-adjusted "ground truth" contours via dice similarity coefficient (DSC) and 95-percentile Hausdorff distance (HD95%). Once contours were approved, adapted RT plans were auto-generated for physician review. Simulated doses to OARs and targets from the adapted plan were compared to that from the original (un-adapted) plan. Results Median on-line ART planning duration in the remote emulated environment was 19 min 34 sec (range: 13 min 10 sec - 31 min 20 sec). Automated key OAR quality was satisfactory overall (98% scored ≥3; 82% ≥4), though mandible was rated lower than others (p < 0.05). Most key OARs and all targets were within 2 mm margin of human-adjusted contours, but a few parotid and spinal cord contours deviated up to 5 mm. Anatomical changes over tx course further increased auto-contour error (p < 0.05, ΔHD95% = 0.77 mm comparing start and end of tx). Further optimizing auto-contoured OAR and target quality could reduce the on-line treatment planning duration by ~5 min and ~4.5 min, respectively. Dosimetrically, adapted plan spared OARs at a rate much greater than random chance compared to the original plan (χ² = 22.3, p << 0.001), while maintaining similar therapeutic dose to treatment target CTV (χ² = 1.14, p > 0.05). In addition, a general decrease in accumulated OAR dose was observed with adaptation. Unsupervised adapted plans where contours were auto-generated without human review still spared OAR at a greater rate than the original plans, suggesting benefits of adaptation can be maintained even with some leniency in contour accuracy. Conclusion Feasibility of a novel, semi-automated on-line ART system for various head and neck (H&N) cancer sites was demonstrated in terms of treatment duration, dosimetric benefits, and automated contour accuracy in a remote emulator environment. Adaptive planning duration was clinically viable at 19 min and 34 sec, but further improvements in automated contour accuracy and performance improvements of plan auto-generation may reduce adaptive planning duration by up to 10 minutes.

Keywords: adaptive radiotherapy; automatic planning; head and neck oncology; image segmentation; online adaptive radiotherapy.

Figure 1. On-line ART workflow for H&N patients.

During initial planning, pre-treatment planning CT is acquired and OAR and target contours are either imported from TPS or manually contoured within the adaptive planning software. On-the-day H&N on-line ART begins with 1. CBCT acquisition. 2. Planning CT is deformably registered to this CBCT (unique to H&N workflow). 3. Influencer is propagated (unique to H&N workflow). 4. Influencer is reviewed, from which an adjusted deformation field is derived. This human-adjusted deformation field is used to propagate. 5. OAR and 6. target contours which are also 7. reviewed. Corrected contours are used to plan and calculate dose from 8a. original (also called scheduled) and 8b. supervised adapted plans. *Influencers are OAR contours with clinical significance, which for H&N sites are the parotids, spinal cord, and mandible. Human modifications to these contours "influence" the CT-to-CBCT deformation vector field. ART: Adaptive radiotherapy; OAR: Organs-at-risk; TPS: Treatment planning software; H&N: Head & neck; CBCT: Cone-beam CT; RT: Radiotherapy.

Figure 2. Overview of on-line ART workflow duration in each module and summary of human intervention outcomes.

ART: Adaptive radiotherapy; OAR: Organs-at-risk; PTV: Planning target volume.

Figure 3. Subjective and objective comparisons of auto-generated influencer and target contours with human-generated contours.

(A) Histogram of overall subjective quality of auto-contoured influencers, rated from 1-5 where 1 is unacceptable delineation of the organ (i.e., majority of the auto-contour is outside the organ, or majority of the organ is outside the auto-contour), 3 is moderate modifications required to accurately represent the organ (i.e., auto-contour moderately over-contours or under-contours the organ), and 5 is perfect delineation of the organ without modification. (B) Breakdown of subjective quality by organ (top) and treatment progress (bottom). Auto-contoured mandible was found to be worse quality subjectively than the other three, and the auto-contour subjective quality dropped at later radiotherapy fractions. (C) Table of overall objective quality of auto-contoured influencers compared to human-contoured counterpart, assessed with dice similarity coefficients (DSC), mean Hausdorff distances, and 95% Hausdorff distances (HD95%). (D) When grouped by treatment progress, auto-contoured target DSC was found to worsen with treatment progress (p = 0.033). (E) DSC and HD95% for each of the four auto-contoured organs were similar, signifying the auto-contouring algorithm did not perform better for any one organ versus others. (F) When grouped by treatment progress, influencer auto-contour HD95% was found to increase near the end of the treatment. This indicates auto-contour quality worsens at later radiotherapy fractions.

Figure 4. Dosimetric comparison of adapted versus original (scheduled) plans.

(A) OAR sparing and target (CTV) coverage for daily adapted plans versus original plan. The table shows the adapted plan met the dose constraints (or clinical goals) for the 300 OARs studied more often than original un-adapted plan (p << 0.001), but not for 36 CTVs studied (p = 0.28). CTV dose difference from the clinical goals is shown on upper right. OAR dose difference from the clinical goals is shown on bottom, grouped by the type of clinical goal (D_mean or D0.03cc). (B) Accumulated dose to OARs and targets from one representative adapted plan for a full 70-Gy prescription H&N RT, expressed as a difference from the original plan. The sum of daily doses from adapted plan is generally lower than that from original plan for both D_mean and D0.03cc clinical goals (left and top right), but is similar for the CTV D99% with <1 Gy difference for the whole treatment (bottom right). Ph. Constr. = pharynx constrictor; Submand. Gl R = right submandibular gland

Figure 5. Dose consequences of the lack of supervision on human-adjusted contours (unsupervised adapted) versus not adapting at all (original), compared to supervised plan.

Figure 6. Various factors that significantly impact the duration of human review of automated influencer and target contours during on-line ART.

(A) Poor subjective auto-contoured influencer quality extends review time by about 300 seconds. (B) Influencer review time increased linearly with mean Hausdorff distances of the four auto-contoured influencers from their respective human-adjusted contours, indicating worse objective measure of auto-contour quality leads to longer review time. (C) About 360 seconds was saved from target review time if target auto-contour was satisfactory to physicians without adjustments. (D) Treatment progress, which decreased auto-contoured influencer quality in our experience, surprisingly did not affect influencer review time.