Heidelberg 2022 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

ST: Fachverband Strahlen- und Medizinphysik

ST 3: Artificial Intelligence in Medicine

ST 3.6: Talk

Tuesday, March 22, 2022, 12:15–12:30, ST-H4

Preparing transformer-based dose predictions: Performance of encoder/decoder structures for CT- and dose-sequence encoding — •Piet Hoffmann, Kevin Kröninger, Armin Lühr, Florian Mentzel, and Jens Weingarten — TU Dortmund

In radiotherapy, fast dose predictions based on CT images are useful as they reduce the need for computing-intensive Monte Carlo simulations and thus can speed up treatment planning. A new approach to these fast dose predictions consists of interpreting the CT to dose conversion as a sequence translation task and making use of a transformer machine learning model.

For this, the CT data is disected perpendicularly to the beam into a sequence of 2D slices, if not already aquired in this direction. Before these slices are fed into the translation architecture it is useful to first encode them to reduce their dimensionaltiy and concentrate the contained information. After translation the data then has to be decoded into a dose prediction slice.

For the whole model to work properly, the structure of such encoder and decoder is important and thus in this talk different approaches are compared with respect to their performance. Properties of the resulting encoded data space, like smooth transitioning between data points and the density distribution of data points, and their potential benefits are discussed.