Computational Patient Models

We are physicists, medical informaticians, data scientists, mathematicians, and radiooncologists working together for the benefit of the patient.

Our research is dedicated to in-silico radiotherapy simulation of cancer patients undergoing curative irradiation treatment. The first challenge for technically precise radiation therapy lies in adapting dose delivery to deformations in the patient’s anatomy induced by motion and physiological changes. The second challenge is to detect subtle changes in longitudinal imaging, which allow to test treatment outcome prediction models.

Our vision is to create a virtual patient model, serving as a proxy for treatment simulation by accumulating the physical dose and predicting the therapeutic effect of the planned treatment. This together empowers treating physicians to tailor the treatment to the specific needs of the patient before irradiation and during the treatemtn course. In-silico model development, advancements in imaging and computer technology, and our passion for unusual solutions are our tools in the fight against cancer.

CBCTart - Data-Driven CBCT Image Quality Improvements for Online Adaptive Radiotherapy

CBCT imaging has become an integral component of photon radiotherapy devices. However, its image quality is restricted in its ability to faacilitate precise patient positioning. The primary objective of the project is to bridge the gap toward achieving image quality comparable to planning CTs, thereby enabling on-couch treatment plan adaptation based solely on CBCTs. 

In this Varian-funded project the team investigates the impact of typical image quality shortcomings of CBCT on the degradation of plan quality. Additionally, generative deep learning models are employed to generate synthetic CTs from the lower-quality CBCT images. 

Team: Goran Stanic(DKFZ), Kristina Giske (DKFZ), Niklas Wahl (DKFZ), Florian Ebert (DKFZ), Oliver Jäkel (DKFZ)

Collaborators: Fabian Weykamp (UKHD, KKE DKFZ), Bálint Kovács (MIC DKFZ)

TVoracle - Towards Expert-Guideline Conformance for Machine-Learning-based Segmentations of Clinical Target Volumes

Patient-tailored contours of target volumes are crucial for radiation treatment planning, significantly influencing the outcome of cancer treatment. Clinical target volume delineation on planning CT scans poses challenges for human experts, proving to be extremely time-consuming and displaying substantial variation between observers. State-of-the-art machine learning algorithms achieve high accuracy on the automatically segmenting anatomical structures, but this success doesn't seamlessly extend to target volumes without additional constraints. The translation of expert guidelines into the machine learning realm has the potential to advance automated target volume delineation, facilitating guideline conformance and clinical use. 

In this HIDSS4Health-funded project the team tackles the implications of supervised learning on data prone to inter-observer variabilities and patient-individual differences utilizing mathematical rules provided by human experts.

Team: Alexandra Walter (KIT, DKFZ), Jakob Kreft(DKFZ), Marc Buckmakowski (DKFZ), Kristina Giske (DKFZ), Oliver Jäkel (DKFZ), Martin Frank (KIT)

Collaborators: Philipp Hoegen-Saßmannshausen (UKHD), Sebastian Adeberg (UKHD)

ALIEN - Generating non-existent Radiotherapy Patient Cousins for FAIR Research

Reseach progress in medicine relies on patient-sensitive data, particularly in radiation therapy, which involves the most diverse and frequent imaging requirements for each patient within its precise image-guided adaptive treatment options. Sharing such rich connected data in compliance with ethical and legal requirements necessitates thorough preparations in long-term projects. Technological short-term projects cannot share in-house data, preventing a direct comparison of competing algorithm designs. Creating a data cohort of non-existent but patient-representative radiotherapy-relevant scans could be made openly accessible for radiotherapy research. 

In this AI Health Innovation Cluster-funded infrastructural project our team scrutinises cutting-edge AI image generator technologies for their capability to mimic medical scans with consistently realistic patient anatomies.

Team: Pedro Joao Rodrigues (DKFZ), Jule Katzenberger (DKFZ), Kristina Giske (DKFZ), Oliver Jäkel (DKFZ), Jens Fleckenstein (UMM)

Collaborators: Martin Niklas (DKFZ)

DeepSPYN - Deep Learning based PseudoCT Generation for MR-only Treatment Planning

Stopping power ratio (SPR) maps are essential for dose deposition calculations in ion beam cancer treatments and are usually estimated from single energy CT (SECT) in clinical practice. Dual-energy CT (DECT), which involves the acquisition of two energy spectra, captures both material-specific information and tissue characterization, showing promise in enhancing patient-tailored SPR map conversion. 

In this BMBF*-funded project our team scrutinises the capability of deep learning models to convert MRI scans into pseudo-CTs or pseudo-SPR maps. 

Team: Ama Katseena Yawson (DKFZ), Nora Wolf (DKFZ), Kristina Giske (DKFZ) 

Collaborators: Thomas Welzel (UKHD), Katharina Seidensaal (UKHD), Stefan Dorsch (DKFZ), Katharina Maria Paul (DKFZ), Habiba Salem (UKHD), Julia Bauer (UKHD)

*ARTEMIS Project (#13GW0436B) in collaboration with UKHD & HIT (Jürgen Debus, Oliver Jäkel)

BionicDIR - Towards Biofidelic Deformable Image Registration of the Skeleton by Kinematically Articulated Multi-Body Physics

Accurate motion estimation between medical scans of different modalities (e.g. CT, MRI) is crucial for adaptive radiation therapy. The widely used intesity-based deformable image registration is limited by the absence of physical constraints during deformation. The KinematicDIR project studies the use of a biomechanical model for biofidelic image registration. For the head and neck area, it focusses on a kinematically articulated skeleton model as the driving motion model. 

In this BMBF*-funded  project our team investigates the best approach to incorporate the body model into the registration task and assesses the accuracy and robustness - but also limitations - of the model-based approach. 

Team: Cornelius Bauer (DKFZ), Sina Thürwächter (DKFZ), Kristina Giske (DKFZ)

Collaborators: Thomas Welzel (UKHD), Katharina Seidensaal (UKHD), Katharina Maria Paul (UKHD)

*ARTEMIS Project (#13GW0436B) in collaboration with UKHD & HIT (Jürgen Debus, Oliver Jäkel)

based on PuppetMaster preceding projects by Hendrik Teske, Kathrin Bartelheimer, Peter Lysakowsky, Stephen Schaumann, Jan Meis