Automatic Workflow validation for H&N RT Treatment

Purpose: Automatic techniques aim to reduce the manual operator work-time and inter-operator variability without compromising the treatment quality. The objective of this study is to evaluate the performance of automatic workflow for head-and-neck (H&N) cancer radiotherapy using a multi-atlas based auto-contouring software and an a priori multicriteria plan optimisation algorithm for three distinct dose prescriptions on a cohort of different H&N patients. Following their validation, a feasibility study to implement an adaptive online approach with CBCT images is carried out. For the treatment optimisation, two different modalities are explored, the adapt to position (ATP) and the adapt to shape (ATS). The latter mode focuses on plan optimisation based on daily patient anatomy, where a dose accumulation analysis can also be performed. Materials and Methods: A research version of the commercial software ADMIRE (Elekta AB, Stockholm, Sweden) for the multi atlas-based segmentation of H&N structures on CT images is used. To create a multi-atlas database, the CT scans of nine patients are selected. Thirty-three anatomical structures and five HU layers (air, bones, tissue, fat, patient) are contoured on each CT dataset with the support of a radiotherapist. The geometric accuracy of two different auto-segmentation methods, STAPLE and Random Forest (RF), is evaluated in terms of Dice similarity coefficient (DSC) and Hausdorff distance (HD), using a leave-one-out cross-validation approach and considering the OAR manually segmented as the gold standard. The mCycle algorithm, implemented in a research version of Monaco TPS (Elekta AB, Stockholm, Sweden) is used for the a priori multicriteria plan calculation. A total of twenty H&N patients are selected for this planning part. For each patient three different dose prescriptions are investigated. Automatically generated plans are obtained using a parameterised wish-list, capable of three different prescriptions. They are compared to manual VMAT plans by assessing differences in planning time, dose delivered to targets and organs at risk (OARs), and calculating the plan quality indexes (PQIs). Through a signed-rank Wilcoxon test the statistical significance of each analysed parameter is assessed. A clinician also performed a blinded score evaluation and comparison of each plan. Deliverability analysis of the plans, with the real assigned prescription, is carried out. Two representative patients are then chosen for the retrospective CBCT adaptive online feasibility analysis. For eleven fractions, an ATS and ATP plan is generated. To assess the differences for the two adaptive modalities the clinical goals for targets and OARs and the percentage passing rate of constraints are explored for each fraction. Signed-rank Wilcoxon test is performed to asses the statistical relevance. An analysis of the timing for the different steps required to produce an online adaptive plan is also carried out to assess its clinical applicability. For the dose accumulation the 35 fractions of one of the previous patients are considered. For each fraction an ATS plan is calculated and then back-projected to the reference CT. The warped dose maps are then normalised and a summation is performed to accumulate the dose distributions. Results: The RF algorithm shows better results compared to the STAPLE and proves to be a reliable tool for automatic delineation. Nevertheless, the RF algorithm does not perform correctly for some structures and further investigations are necessary to improve its performance. RF algorithm proves to be a valuable and time-efficient solution compared to manual contouring by operators. The mCycle auto-planning significantly reduces planning time while maintaining clinical acceptability, enhancing organ at risk sparing and preserving a good target coverage. The radiotherapist deems all plans clinically acceptable, and in the majority of cases (83\%), the automatic plan scores equal to or higher than the manual plan. Automatic plans reach superior results for the cochlea, eyes, lachrymal glands, larynx, lens, lips, oral cavity, constrictor muscles, and thyroid, also confirmed by the PQI_OARs. Number of MUs, total delivery time and modulation degree are significantly higher for automatic plans compared to manual ones. Comparing ATS and ATP modes in adaptive radiotherapy, ATS exhibits superior outcomes. The targets are the most frequently failing the dose constraint for ATP. The time efficiency is crucial in online adaptive approaches: using the ATS mode the planning time takes around 14 minutes and approximately 20 minutes for the entire treatment. For the dose accumulation no relevant differences are observed between the values of the original treatment plan and the plan obtained through dose accumulation of ATS plans. Conclusions: This study contributes to the advancement of automatic and adaptive radiotherapy, demonstrating the potential of an automated workflow in H&N treatments. The successful validation of auto-contouring and auto-planning software, combined with preliminary findings on online adaptive, underscores the significance of exploiting technology to optimise treatment and improve care for radiotherapy patients.