Computational medical physics: Target definition and treatment planning
Figure: Three selected patient’s original CT, PET, MRI (T2, T1) images, deep learning predicted contours are compared from different image combinations. (B): GTV size is 5.6 cm3 , a case with small tumor size. (C): GTV size is 54.1 cm3 , a large tumor case. (D): GTV size is 35.5 cm3 , a case with planning CT artifacts.
We use computational methods such as machine learning, data mining and inverse optimisation, to challenge existing paradigms in radiotherapy for development of novel treatment planning methods. We use advanced analysis of medical images including CT, PET and MR scans, in large patient cohorts, with focus on head and neck cancer and breast cancer. We specifically address the problems that give rise to the largest sources of uncertainty in radiotherapy planning; namely target definition, dose optimisation, and anatomical changes during treatment.
Target delineation is a time-consuming process, which is also prone to large uncertainties and subjective assessment by the delineating clinician. We use deep learning to automatically segment the gross tumour volume in multimodality images, for use in interactive delineation tools. By introducing a deep learning generated segmentation in an interactive user interface, we aim to speed up the delineation process while at the same time reducing the uncertainty related to the delineation for a more accurate treatment.
We investigate the implementation of risk models for tumour spread, which can be used for probabilistic optimisation of dose distribution around the tumour. This approach will give the potential to reduce high radiation doses to volumes of tissue where the risk of finding cancer cells is low, and hence reduce the risk of side effects.
Other projects include individualized robust optimization to address anatomical changes during treatment, data mining for modelling patient individual risk factors, and development of automated treatment planning methods.
Overall, we aim to enhance tailoring of radiotherapy to the individual patient through adaptation to patient specific risk factors from medical images. Moreover, we expect that the introduction of automated and computational methods will lead to an overall decreased level of uncertainty in the treatment and increased treatment quality at both individual and population basis.
Head and neck cancer:
- Deep learning for automated tumour segmentation
- Computational target definition with voxel-based risk modelling
- Individualised probabilistic treatment planning
- Robust treatment planning for proton therapy
- Data mining for cardiac toxicity risk factors and modelling
- Data mining for mapping of delineation and treatment planning practices
- Automated delineation and treatment planning