ESTRO 2024 - Abstract Book

S4514

Physics - Machine learning models and clinical applications

ESTRO 2024

2006

Proffered Paper

Personalized delineation sensitivity envelopes - where to adjust organ at risk delineations?

Erik van der Bijl 1 , Tomas M. Janssen 2 , Steven F. Petit 3

1 Radboud University Medical Center, Radiation Oncology, Nijmegen, Netherlands. 2 Netherlands Cancer Institute, Radiation Oncology, Amsterdam, Netherlands. 3 University Medical Center Rotterdam, Department of Radiotherapy, Erasmus MC Cancer Institute, Rotterdam, Netherlands

Purpose/Objective:

Automatic segmentation via deep learning or contour propagation facilitates faster workflows, but radiation oncologists / therapists still need to review and/or modify automatically generated segmentations. This process often still is time-consuming and time could be saved by only looking at the parts that influence the treatment plan most or significantly. For example, delineation adjustments of sections of a dose limiting organs-at-risk (OAR) close to a target probably affect treatment plans more than delineation errors farther away. However, the effect of a delineation error on the treatment plan is not straightforward to determine prior to treatment planning, since it depends on the interplay between the location of the error, patient anatomy, the dose constraint under consideration and on the trade-off in the treatment plan itself. The goal of the current work is to give a quantitative estimate of the sensitivity of the treatment plan to the local adjustment of OAR segmentations. To this end we introduce the concept of patient specific delineation uncertainty envelopes. These envelopes are defined such that any edits within the envelope do not influence the treatment plan metrics within a pre-defined threshold. The envelopes thus guide the user to decide when to stop adapting automatically delineated contours. For this work we use the principle that the entire treatment planning Pareto front (PF) for prostate cancer patients can be accurately predicted based solely on the patient anatomy, i.e. contours of tumor and OAR[1] and their overlap volume histograms (OVH). This PF considered was three dimensional, describing the trade-off between target homogeneity, high dose conformity and rectum mean dose, for a fixed target coverage. For each patient from a group of 110 prostate cancer patients coming from a public Prostate Treatment Planning QA dataset [2] uncertainty envelops were created. To create the envelopes the rectum was divided into 10 sections corresponding to the rectum superior (S) and Inferior(I) to the PTV and eight sections at the level of the PTV. This was subdivided in left(L), right(R), anterior(A) and posterior(P) sections of the rectum. These 4 were divided into upper( ↑ ) and lower ( ↓ ) halves, see Figure 1. Subsequently, for each section we created local variations of the rectum, by contracting or expanding only this section from -10 to +10 mm, in steps of 0.2mm, while fixing the other sections. For each of the perturbations of the contours the associated OVH with the PTV was calculated as input for PF prediction model. Sensitivity envelopes were created by selecting, per section, the maximal expansion and contraction that led to a maximum deviation of 0.5 Gy of the three asymptotes that span the PF for the unperturbed delineation. Material/Methods: