Pre Treatment Patient QA results prediction using deep learning, based on bayesian classification.
Résumé
Purpose or Objective
The evolution of radiation therapy treatments sees the increasing use of VMAT techniques, with often small and very irregular fields, in order to generate highly modulated treatments. Therefore systematic pre treatment QA for each beam is needed to ensure that calculated and measured dose are within tolerance, The objective of this study is to use metrics to quantify the complexity of a
treatment plan in order to assess the relevance of the QA.
Material and Methods
Sixteen complexity metrics have been have been identified as relevant: fifteen from literature (PIMV, AI, MCS, MI…) and one novel (wavelet transform of the fluence map). They can be classified into two categories: the metrics based on geometrical aspects (shapes of the fields, opening of the leaves), and based on fluence. Their calculation use dicom RT files generates by TPS (treatment
planning systems). We have used machine learning techniques to create a model linking the Patient QA results and to complexity metrics. Two types of models have been created (pass/fail
prediction, based on nonparametric bayesian method) and gamma-index features (passing rate, mean gamma, max gamma, based on regression). Prediction uncertainty has been implemented with a prediction model assessment using the leave-one-out method LOOCV (logarithmic, zero/one and R² scores) and the leave-pair-out method LPOCV (ROC curves and AUC). We have built a prediction model, at APHM – La Timone, France, based on the following equipment ( Elekta Synergy, Beam Modulator, Delta4, Pinnacle3, and different tumor sites (pelvis, prostate, H&N, thorax, brain, …)). 445 VMAT plans with 615 arcs have been used in the learning phase. The clinical gamma criteria used was 3% local dose/3mm/ treshold 20%. For the test phase, 92 plans with 146 arcs have been used.
Results
ROC Analysis has been performed on the learning database (Area Under The Curve = 0,85, Zero-one = 0,89), that is linked with the specific equipment used ( Elekta Synergy, Beam Modulator, Delta4, Pinnacle3, and different tumor sites (pelvis, prostate, H&N, thorax, brain, …)), and the specific acceptance criteria used. (3% local dose/ 3mm/treshold 20%).
Conclusion
We have shown that with this deep learning method, we could reduce the number of pre treatment patient QA by 45% (True OK/KO) and have a good efficiency of the global prediction of 89% (True OK/KO/TEST). The 11% left being good plans that will be tested. Further investigations in other radiation therapy centers are ongoing, and the Gamma values will also being studied, in order to give additional analysis tools to reduce the time spent for pre treatment patient QA.
Mots clés
radioactivity
ionizing radiation
radiotherapy
modulated treatment
dosimetry
dose simulation
dose measurement
modelling
simulation
treatment planning system
machine learning
artificial intelligence
leave-one-out method
leave-pair-out method
tumor
cancer
learning database
deep learning
online learning