ESTRO 2024 - Abstract Book

S3946

Physics - Image acquisition and processing

ESTRO 2024

1 Friedrich-Alexander-Universität Erlangen-Nürnberg, Pattern Recognition Lab, Erlangen, Germany. 2 Siemens Healthcare GmbH, Cancer Therapy, Forchheim, Germany. 3 Siemens Healthcare GmbH, Computed Tomography, Forchheim, Germany. 4 Friedrich-Alexander-Universität Erlangen-Nürnberg, Department of Radiation Oncology, Erlangen, Germany. 5 Universitätsklinikum Erlangen, Comprehensive Cancer Center Erlangen-EMN (CCC ER-EMN), Erlangen, Germany

Purpose/Objective:

Time-resolved, four-dimensional computed tomography (4DCT) imaging is an essential tool for radiation therapy treatment planning and delivery for moving tumors. One challenge of 4DCT is the occurrence of artifacts, which can limit the use of the images for radiotherapy planning. Common artifacts are so-called stacking or stair-step artifacts, that appear when the patient's breathing varies strongly between the multiple scan positions that are needed to generate a volume of the whole lung. Thus, breathing irregularity can be an indicator of image quality. Most artifact avoidance or reduction techniques are applied during or after the 4DCT scan [1, 2]. In contrast, our approach could allow to prevent unusable, artifact-affected images before any radiation is emitted. By enabling a risk assessment based on the patient-specific breathing pattern, countermeasures can be taken in advance in the form of the selection of a different scan mode, patient specific scan parametrization or breathing coaching. The same breathing irregularity analysis methods could also be applied directly after the scan to immediately obtain an indication of the probability of a rescan. Additionally, it could assist in the rescan decision by indicating areas where irregularities are likely because of irregular breathing patterns.

Material/Methods:

Based on respiratory curves, acquired from patients receiving 4DCT imaging, we train an algorithm that predicts whether the image quality is sufficient for planning radiotherapy or if the experts would consider a rescan.

As a training dataset we use data from a previous study, in which five medical physicists and five physicians evaluated the image quality of 56 4DCT sequence scans by blind assessment [1]. All scans were categorized by whether they contained artifacts that may possibly require a rescan (Likert ratings 1-3) or minor to no artifacts (Likert ratings 4-5). Scans were acquired on a SOMATOM go.Open Pro scanner (Siemens Healthineers, Forchheim, Germany) while tracking respiratory motion using the RGSC system (Varian Medical Systems, Palo Alto, CA, USA). A large variety of breathing signal characteristics might be of interest for their influence on image quality, since in combination they are able to represent common irregularities like variations in amplitude or frequency, baseline drift, and breathing pauses. We employed a feature generator known as tsfresh [3], to calculate over 700 features. These included absolute energy, autocorrelation, Fourier coefficients, linear trend, among others. These features are provided as input to four machine learning classification algorithms (Decision Trees, Naïve Bayes, RandomForest, XGBoost). The optimal hyperparameters for the algorithms were chosen by using grid search with five-fold cross-validation.

Results: