144. Machine learning outcome prediction model-based radiotherapy treatment plan optimization using the open-source toolkit pyanno4rt
Invited abstract in session WC-4: Optimization in regression, classification and learning I, stream Optimization in regression, classification and learning.
Wednesday, 10:05 - 11:20Room: M:M
Authors (first author is the speaker)
| 1. | Tim Ortkamp
|
| Karlsruhe Institute of Technology (KIT) | |
| 2. | Martin Frank
|
| Karlsruhe Institute of Technology (KIT) | |
| 3. | Oliver Jäkel
|
| German Cancer Research Center (DKFZ) | |
| 4. | Niklas Wahl
|
| German Cancer Research Center (DKFZ) |
Abstract
Inverse radiotherapy treatment plan optimization problems can be described as continuous, multi-objective, nonlinear, potentially nonconvex and large-scale (up to hundreds of thousands of variables). Conventionally, these problems are formed by translating clinical criteria into a set of mathematical objectives, which rely on empirical dose prescription and tolerance parameters rather than directly optimizing for the treatment outcome as quantified by normal tissue complication probability (NTCP) and tumor control probability (TCP). Classical outcome models like the Lyman-Kutcher-Burman model are technically feasible for optimization, but questionable with regards to accuracy and usability. Our contribution therefore lies in the development of a multivariate machine learning (ML) model-based optimization framework. To this end, we implemented pyanno4rt, an open-source Python package featuring different optimization methods using zero-, first- and second-order solvers, along with various strategies for model integration, efficient calculation via JIT compilation, and automatic differentiation. We internally fitted three ML models (logistic regression, neural networks, support vector machine) on two head-and-neck (N)TCP datasets, then optimized and evaluated conventional, radiobiological and ML model-based treatment plans for an example patient. Our results show that applying ML models can yield acceptable dose distributions with improved outcome predictions.
Keywords
- Data driven optimization
- Large- and Huge-scale optimization
- SS - Multi-Criteria Optimization and Real-World Applications
Status: accepted
Back to the list of papers