Author: Martin Frank, Oliver Jäkel, Niklas Wahl 👨🔬
Affiliation: Department of Medical Physics in Radiation Oncology, German Cancer Research Center (DKFZ), Karlsruhe Institute of Technology (KIT) 🌍
Purpose: Machine learning (ML) models on normal tissue complication and tumor control probability ((N)TCP) exploiting e.g. dosiomic and radiomic features are playing an increasingly important role in radiotherapy treatment planning and optimization. Building on previous work on the technical feasibility of integrating these models into IMPT/IMRT optimization schemes, we intend to show that our framework can also generalize well to multiple patients with different tumor sites and anatomies.
Methods: We optimized 28 conventional and outcome model-based treatment plans for seven different head-and-neck cancer patients using our self-developed open-source Python module "pyanno4rt", which combines treatment plan optimization with ML outcome modeling. In this context, we fitted three ML outcome models (logistic regression, neural network, support vector machine) for grade 2+ long-term xerostomia probability prediction, using data with up to 34 features over a cohort of 153 head-and-neck cancer patients, and embedded the models within the inverse planning problem using feasible model transformations and end-to-end differentiation. Finally, we analyzed the resulting iterative NTCP curves and dosimetric plan indicators for each patient.
Results: Our framework generalizes well to multiple patients, translating into reduced NTCP levels for all 28 plans (mean NTCP reductions: 3.8 % (logistic regression), 11.5 % (neural network), 50.9 % (support vector machine)). Target coverage is not compromised by applying the models, and only slight variations in organ-at-risk sparing are evident due to the different learned feature importances of the models.
Conclusion: High-dimensional multivariate ML outcome models can not only be integrated in a technically feasible way, but can also be applied to multiple patients with different characteristics. Future work should address the quality of the outcome models and the integration of clinically validated models.