Speaker
Description
Introduction
Delivering Carbon-ion radiotherapy (CIRT) at consistently high accuracies throughout a multi-week treatment remains challenging in clinical practice. Even though the dose conformity of CIRT is exceptionally high, it always relates to the last known factual state of the patient’s anatomy, which gets increasingly uncertain the more time has passed since the Computed Tomography (CT) imaging for dose planning was acquired. Since current treatment guidelines often include at most weekly control imaging throughout a multi-week therapy, our goal is to update the knowledge about the state of the patient anatomy at daily intervals. Detecting clinically significant changes between treatment fractions can support clinicians in authorizing additional imaging to adapt the treatment plan when necessary.
Our group’s approach to obtaining additional evidence about anatomical changes is to detect nuclear fragments, emitted during irradiation, with Timepix3 detectors and reconstructing their point of origin [1]. The spatial distribution of secondary radiation has been shown to carry significant information about the treated region and we develop methods capable of exploiting it to perform in vivo monitoring. The goal is to detect and localize anatomical changes between treatment days and extract their 3D volume and shape information for further evaluation. This will support reducing the reliance on conventional imaging modalities such as CT, which involves additional resources and radiation exposure for the patient.
Materials and Methods
Our approach is based on our recently published method for detecting the depth of anatomical changes along the beam axis by exploiting joint frequency band variations of projection differences [2]. At the center of our method, a distinct correlation matrix enables identifying local similarities and consequently distinguishing between the target region, i.e. the locations of the changes, and the background.
We extract the required 3D information about potential anatomical changes by performing the same depth (z) analysis consecutively for all nuclear fragment subsets contained in a moving window in the transverse plane (x,y). Joining the window locations with the detected depth yields the 3D reconstructed locations of anatomical change. This splitting of the original set of reconstructed nuclear fragment origins into smaller subsets for analysis has the consequence that each subset exhibits a statistically less stable distribution. Given the already challenging signal-to-noise ratio when comparing measurements from different treatment fractions, reducing the available number of fragment origins makes robust detection and localization increasingly challenging.
We therefore trained a deep-learning model with the aim of determining the depth of anatomical changes based on classifying correlation matrices to support our reconstruction pipeline. The method treats the depth approximation as an image classification problem, exploiting high-dimensional information extracted from this already enriched data source and compressing it further into a neural network model (ResNet, RegNet or ConvNeX architectures) which performs its own depth approximation. Both results are subsequently joined into a common 3D reconstruction.
Results
We validated our 3D reconstructions on measurements acquired during dedicated experiments with Polymethyl Methacrylate (PMMA) head sized phantoms conducted at the Heidelberg Ion Beam Therapy Center (HIT). Our experimental setups contained coin-shaped air cavities of 20 mm diameter placed at varying depths and lateral locations of the phantom. These controlled setups allowed a detailed qualitative and quantitative assessment of the method in terms of reconstruction loss versus the known 3D setups. We report the results at varying depths, with a mean absolute error over the region of interest starting from <3 mm and Jaccard Index (intersection over union, or ‘IoU’) values peaking at >0.8 within the region of interest when considering the exact 3D location of the cavity. We show that our framework is capable of highly accurate 3D reconstruction of anatomical changes, including retaining the principal location and shape of the change region.
Conclusion
Our results show that we can achieve full 3D reconstruction of anatomical changes based only on the information carried by nuclear fragments emerging from the phantom. We have shown that despite the small number of available data sets for training, it is possible to extract sufficiently localized information to successfully train modern deep-learning models to serve as supporting methods and surrogates useful in stabilizing statistical variations. Mapping the exact locations to those of the actual 3D setup, we can see that a delineation of coin-shaped change regions is not only feasible but performs well at reasonably shallow depths. We therefore have reason to conclude that our method constitutes a useful tool which can support clinical decision making, reducing the need for additional conventional imaging.
Literature
[1] L. Kelleter et al., Sci. Rep. 14 (2024), 15452, doi: 10.1038/s41598-024-66266-9
[2] P. Schlegel et al., Phys. Med. Biol. 70(24) (2025), 245009, doi: 10.1088/1361-6560/ae22bb
The authors acknowledge funding by Helmholtz Information & Data Science School for Health (HIDSS4Health).