Principal component analysis modeling of Head‐and‐Neck anatomy using daily Cone Beam‐CT images

Abstract

Purpose To model Head‐and‐Neck anatomy from daily Cone Beam‐CT (CBCT) images over the course of fractionated radiotherapy using principal component analysis (PCA). Methods and materials Eighteen oropharyngeal Head‐and‐Neck cancer patients, treated with volumetric modulated arc therapy (VMAT), were included in this retrospective study. Normal organs, including the parotid and submandibular glands, mandible, pharyngeal constrictor muscles (PCMs), and spinal cord were contoured using daily CBCT image datasets. PCA models for each organ were developed for individual patients (IP) and the entire patient cohort/population (PP). The first 10 principal components (PCs) were extracted for all models. Analysis included cumulative and individual PCs for each organ and patient, as well as the aggregate organ/patient population; comparisons were made using the root‐mean‐square (RMS) of the percentage predicted spatial displacement for each PC. Results Overall, spatial displacement prediction was achieved at the 95% confidence level (CL) for the first three to four PCs for all organs, based on IP models. For PP models, the first four PCs predicted spatial displacement at the 80%–89% CL. Differences in percentage predicted spatial displacement between mean IP models for each organ ranged from 2.8% ± 1.8% (1st PC) to 0.6% ± 0.4% (4th PC). Differences in percentage predicted spatial displacement between IP models vs the mean IP model for each organ based on the 1st PC were <12.9% ± 6.9% for all organs. Differences in percentage predicted spatial displacement between IP and PP models based on all organs and patients for the 1st and 2nd PC were <11.7% ± 2.2%. Conclusion Tissue changes during fractionated radiotherapy observed on daily CBCT in patients with Head‐and‐Neck cancers, were modeled using PCA. In general, spatial displacement for organs‐at‐risk was predicted for the first 4 principal components at the 95% confidence levels (CL), for individual patient (IP) models, and at the 80%–89% CL for population‐based patient (PP) models. The IP and PP models were most predictive of changes in glandular organs and pharyngeal constrictor muscles, respectively.

On the complexity of helical tomotherapy treatment plans

“…21 Principal component analysis is a multivariate technique widely used in dataset dimensionality reduction to increase interpretability while preserving most of the initial information. [22][23][24] . For that, PCA finds a new set of uncorrelated variables (principal components, PCs) that result from linear combinations of the original ones and that successively maximize variance.…”

“…8,9 Bresciani et al, 8 using 384 HT plans of multiple treatment sites, found no strong correlations between some of these factors and the results of pretreatment QA verification. Binny et al 9 have used multiple statistical process control methods on a set of head and neck (28), pelvic (19) and brain (23) plans, to define lower and upper limits for planning parameters, like the modulation factor, gantry period, and couch speed, based on acceptable pretreatment QA results. The established ranges were specific to each treatment site and contributed to improve the treatment efficiency at their institution.…”

Evaluating principal component analysis models for representing anatomical changes in head and neck radiotherapy

“…Previous studies have investigated the generation of PCA models in H&N cancer patients and their evaluation. For example, Tsiamas et al [15] , assessed the number of components needed to model the spatial displacements for specific organs, using data from 18 patients to create both individual and population-based models. They focused on comparing the relative variance of the different PCA components between models.…”

“…using robust or probabilistic planning [5] . Principal component analysis (PCA) is widely used for creating such models, for instance in lung [6] , [7] , [8] , [9] , prostate [10] , [11] , [12] , [13] , [14] , cervix [5] and head and neck (H&N) [15] , [16] . The usefulness of such models depends on their ability to accurately simulate future changes in the patient.…”

A probability model for anatomical robust optimisation in head and neck cancer proton therapy

“…Anatomical models that simulate the possible geometric variations from a population of patient data remove the requirement of multiple scanning and the dependence of CT images acquired during the treatment by taking the predicted images into aRO. Several mathematical models have been proposed to account for anatomical changes (Yu et al 2016, Tsiamas et al 2018. Yu et al used an anatomical model for deformable image registration (DIR) evaluation.…”