University of Twente Student Theses

Login

Automated robust planning for intensity modulated proton therapy (IMPT) in oropharyngeal cancer patients using machine learning

Huiskes, M. (2020) Automated robust planning for intensity modulated proton therapy (IMPT) in oropharyngeal cancer patients using machine learning.

[img] PDF
3MB
Abstract:Background: Radiotherapy treatment planning is a complex and time-consuming process. In intensity modulated proton therapy (IMPT) treatment planning steep in-field dose gradients are used, which allows for dosimetric benefits, but making IMPT also sensitive for density- and set-up errors. Robust planning methods have been developed to account for these errors, but robust treatment planning is even more time consuming. In this study, we combined a robust dose mimicking optimization with a machine learning based dose prediction (machine learning optimization (MLO)) to automatically generate robust IMPT plans. We aimed to automatically generate robust IMPT plans for oropharyngeal cancer patients with similar quality as clinically available plans. Methods and materials: A total of 79 robust IMPT plans of oropharyngeal cancer patients were included. Dose distributions, contours and CT image features of 66 patients were used to train a model to predict dose distributions for novel patients. The target coverage during training was based on primary and elective clinical target volume (CTV). Dose prediction was based on a random forest model. Hence, the predicted dose was converted into a deliverable plan using robust voxel-wise dose mimicking optimization, including 21 perturbed dose scenarios with 3 mm isocenter shifts and ±3% density uncertainty. IMPT plans from 8 patients were used for (cross)-validation, and subsequently testing was performed with 9 patients. Targets were assessed in terms of robust target coverage (voxel-wise minimum D98>94%), conformity and homogeneity indices. Organ at risk (OAR) dose was assessed by Dmax, Dmean and (sum) normal tissue complication probability (NTCP)s. All parameters were compared to the clinical plan. Results: Robust target coverage was achieved in 8/9 MLO plans (89%) and voxel-wise minimum dose (D98) was statistically significant higher in the high dose CTV compared to the clinical plans. The sum NTCP was lower or did not increase >2% compared to the clinical plans in 8 of the 9 MLO plans (89%) for xerostomia and in 5 of the 9 MLO plans (56%) for dysphagia. The Dmax constraints to the brain, brainstem, spinal cord and eyes were not exceeded in the MLO plans and comparable to the clinical plan. The mean dose to the parotids, right submandibular and pharyngeal constrictor muscles (PCMs) were on average comparable to the clinical plans, i.e. differences were <1Gy. The average Dmean in the oral cavity was statistically significant higher in the MLO plans (3402±1504 cGy) compared to the clinical plans (3102±1515 cGy) with a p-value of 0.008. Conclusion: The MLO algorithm was able to generate robust IMPT plans for oropharyngeal cancer patients with clinically acceptable robust target coverage, and with similar dose to most OARs as compared to the clinical plans. Further reduction of the dose to specific OARs and improvement of the plan consistency between patients have to be investigated.
Item Type:Essay (Master)
Clients:
University Medical Center Groningen, Groningen, The Netherlands
Faculty:TNW: Science and Technology
Subject:33 physics, 44 medicine, 54 computer science
Programme:Technical Medicine MSc (60033)
Link to this item:https://purl.utwente.nl/essays/80822
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page