Click here to

Session: Treatment Planning: Planning Automation and Assessment [Return to Session]

Deep Learning-Based Dosimetric Plan Quality Assessment for Prostate Radiotherapy

Y Nomura*, C Huang, Y Yang, N Kovalchuk, M Surucu, M Buyyounouski, L Xing, Stanford University, Stanford, CA


SU-F-TRACK 5-6 (Sunday, 7/25/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Purpose: This study developed a novel dosimetric plan quality evaluation method for prostate radiotherapy using a multivariate deep neural network (DNN).

Methods: First, 60 prostate CT images with segmented contours of planning target volume (PTV), clinical target volume (CTV) and organs at risk (OARs) were collected. These were divided into 48 patients for training of a DNN model, 6 for validation, and 6 for testing, respectively. Second, intensity-modulated radiation therapy (IMRT) dose distributions were calculated for each patient from seven equally spaced photon beams. Dose-volume histograms (DVHs) of PTV, CTV and OARs and homogeneity index (HI), conformity number (CN), dose coverage index (DCI) and gradient index (GI) of PTV were calculated for each dose distribution. Third, reference quality of the dose distributions was manually assessed by a treatment planner who had access to volumetric CT images, dose distributions and DVHs. Moreover, an in-house plan quality metric (PQM) technique was applied to calculate analytical dosimetric plan quality values. Finally, a multivariate DNN model was generated to predict PQM or the manual plan quality by using HI, CN, DCI, GI and DVHs as inputs. For evaluation, the DNN-calculated qualities of test data were compared with actual plan quality values.

Results: The DNN model estimates PQM and the manual quality values accurately. Averaged relative absolute error and its standard deviation over all test data were less than 4.00% for both PQM and the manual plan quality values. Computation time for calculating one quality value is 0.03 seconds with a single GPU.

Conclusion: A novel dosimetric plan quality assessment method was established using a multivariate DNN. This technique can predict, as a “virtual” planner, plan qualities calculated from multiple assessment methods with high accuracy. This method will be useful for DNN-assisted treatment planning to improve treatment plans and reduce inter-planner quality variation.



    Radiation Therapy, Statistical Analysis, Dose Volume Histograms


    TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

    Contact Email