Subproject 3: AI-driven methods for dose accumulation
The objective of subproject S3, carried out at the LMU University Hospital Munich, is to develop a novel workflow for dose accumulation using deep learning, tailored to lung cancer patients treated with MR-guided online adaptive radiotherapy (Figure 1).
To this aim, deformation vector fields (DVFs) are required to sum up the dose distributions of the individual treatment fractions on the reference planning anatomy. Since the generation of DVFs is generally improved by providing additional constraints, such as organ boundaries, subproject S3 aims at employing automated segmentation of the thorax region to guide the deformation.
The first objective of S3 will thus be to establish an MR-based deep learning auto-segmentation framework tailored to the lung region and capable of segmenting the lung, heart, esophagus, spinal cord as well as the tumor itself. In a second step, a deep convolutional neural network will be trained to output DVFs directly from a given pair of input images (fraction and planning, see Figure 1). Finally, the DVFs will be used to compute accumulated doses and compared to a state-of-the-art non-AI accumulation workflow.
The accumulated doses will serve as a basis for analysis in terms of patient outcome modelling as suggested in subproject S5 and will also be used in subprojects S6 and S4.
Figure 1. Envisaged AI dose accumulation workflow. Fraction and planning images will serve as input to determine a deformation vector field which is then used to warp the fraction dose to the planning anatomy for dose accumulation.
© LMU
Contact
Subproject Leaders:
Prof. Dr. Guillaume Landry
PD Dr. Christopher Kurz