Machine Learning Based Automated Planning Technique For Nasopharyngeal Carcinoma Radiotherapy

This study aims to develop an automated intensity-modulated radiotherapy(IMRT)technique for nasopharyngeal carcinoma(NPC)based on two types of robust machine learning models using overlap volume histogram(OVH).A total of 115 NPC patients treated with definitive step-and-shoot IMRT in Fujian Cancer Hospital were selected to build a dataset in our study.For each patient,seven target volumes(GTV_T_P,CTV1_P,CTV2_P,GTV_NL/R_P and CTV_NL/R_P)and ten organs at risk(bilateral parotids,brainstem,spinal cord,bilateral optic lens,bilateral optic nerves,pituitary and optic chiasm)were delineated on the computed tomography(CT)images.Before modeling with machine learning,OVHs were derived from CT images for each dataset patient and so were the optimization objectives from the manual high-quality IMRT plan of the corresponding patient.Neural network(NN)and k-nearest neighbour(kNN)were modeled using the OVHs as input and the objectives as output.Another twenty-five NPC patients were selected as a test set and their OVHs were also used as input for the two trained models.Two batches of twenty-five sets of patient-specific objectives were derived based on NN model and kNN model,respectively.Then their IMRT plans were automatically generated with only one-shot optimization using Perl-based scripts in Pinnacle~3 treatment planning system.The corresponding 25 manual IMRT plans were generated by an experienced medical physicist in our institute.Finally,comparison on plan quality among the NN-based,kNN-based and manual plans was carried out by an experienced oncologist.Planning duration and dosimetric parameters were also compared using paired-sample t test among the three groups of plans.Compared to the manual plans,20 NN-based plans performed comparable while 19 kNN-based performed comparable.The three types of plans had similar conformity index in GTV_T_P,GTV_NL/R_P and CTV_NL/R_P(P>0.05).Compared to the manual plans,NN-based plans had significantly lower maximum dose in right optic lens(4.70±0.72 Gy vs.5.02±0.70 Gy,P=0.001).The two types of automated plans had significantly lower maximum dose than the manual plans in the bilateral optic nerves(P<0.05).The mean planning duration was 9.73±1.80 min,10.18±2.17 min,and 57.12±6.35 min for generating the NN-based plans,kNN-based plans,manual plans,respectively.Compared to the manual plans,the automated plans are more efficient because of significantly decreased planning time and better sparing of organs at risk.Moreover,the two types of automated plans with only one-shot optimization can both rival the manual ones,indicating that the automated IMRT technique for NPC developed on NN-based and kNN-based model is available.