DIR-based models to predict weekly anatomical changes in head and neck cancer proton therapy

Abstract

Purpose: We proposed two anatomical models for head and neck patients to predict anatomical changes during the course of radiotherapy. Methods: Deformable Image Registration was used to build two anatomical models: 1) The average model (AM) simulated systematic progression changes across the patient cohort; 2) The refined individual model (RIM) used a patient's CT images acquired during treatment to update the prediction for each individual patient. Planning CTs and weekly CTs were used from 20 nasopharynx patients. This dataset included 15 training patients and 5 test patients. For each test patient, a spot scanning proton plan was created. Models were evaluated using CT number difference, contours, proton spot location deviation and gamma index. Results: If no model was used, the CT number difference between the planning CT and the repeat CT at week 6 of treatment was on average 128.9 HU over the test population. This can be reduced to 115.5 HU using the AM, and to 110.5 HU using the RIM3 (RIM, updated at week 3). When the predicted contours from the models were used, the average mean surface distance of parotid glands can be reduced from 1.98 mm (no model) to 1.16 mm (AM) and 1.19 mm (RIM3) at week 6. Using proton spot range, the average anatomical uncertainty over the test population reduced from 4.47±1.23 mm (no model) to 2.41±1.12 mm (AM), and 1.89±0.96 mm (RIM3). Based on the gamma analysis, the average gamma index over the test patients was improved from 93.87±2.48 % (no model) to 96.16±1.84 % (RIM3) at week 6. Conclusions: The AM and the RIM both demonstrated the ability to predict anatomical changes during the treatment. The RIM can gradually refine the prediction of anatomical changes based on the AM. The proton beam spots provided an accurate and effective way for uncertainty evaluation.

Pre‐treatment analysis of non‐rigid variations can assist robust intensity‐modulated proton therapy plan selection for head and neck patients

“…To apply the deformations between groups of subjects, we need to project the SVFs into the atlas space, in which all the SVFs have the same position and resolution. The atlas was obtained from a group‐wise registration which spatially normalized a cohort of patients [1] 16,31 . In the procedure of the projection, the planning CT (pCT) of each patient was the reference geometry, and the CT acquired during the first treatment week (CT t ) was registered to the pCT to produce $$\pmb {v}_{p\rightarrow t}$$, where p stands for pCT and t stands for the week (in this case $$t=1$$) when the weekly CT acquired.…”

“…This reactive approach to adaptive therapy poses workflow challenges for the busy clinical practice. To mitigate time delay during the offline adaptive process, the use of anatomical modeling was suggested 16,17 . Anatomical models can accurately predict the patients' progressive changes and can therefore be used to create adaptive plans in advance, which can be applied as soon as the adaption threshold is reached.…”

“…To mitigate time delay during the offline adaptive process, the use of anatomical modeling was suggested. 16 , 17 Anatomical models can accurately predict the patients' progressive changes and can therefore be used to create adaptive plans in advance, which can be applied as soon as the adaption threshold is reached. Online adaption is intended for same‐day application.…”

Improving workflow for adaptive proton therapy with predictive anatomical modelling: A proof of concept

“…Details of the mathematical formalism of the anatomical models alongside a validation of the models' predictive power are given in Zhang et al [17], and summarised here. To build the model, we randomly selected a patient from the cohort and applied a leaveone-out cross-validation strategy to obtain the average deformation of our training population (n = 19) per week (average model).…”

“…The predictive ability of our model has been validated based on CT numbers, contours, proton spot location deviations and dose distribution in [17]. Compared with no model, in which predicted images were replaced by planning CT, the predictive model reduced the average CT number difference between predicted CTs and real CTs at week 3 by 18.8 HU with approximately 2 million voxels analysed.…”