Abstract:Normal liver dose is of paramount in liver SBRT due to the risk of radiation induced liver disease (RILD). In our clinic, the generalized equivalent uniform dose (gEUD) of normal liver (liver-GTV) is used to determine if a prescription (Rx) needs to be deescalated (e.g., from 50 to 45 Gy) to limit the risk of RILD. To estimate the normal liver gEUD, a planner often needs to create a preliminary plan with full Rx. The goal of this work was to build a machine learning model using artificial neural network trained with our prior SBRT patients to predict normal liver gEUD purely using the geometric properties of liver and PTV, to eliminate the need for the preliminary plan. The initial model was trained by 40 consecutive liver SBRT patients treated in our clinic from December 2014 to February 2018 meeting the following criteria: (1) 50 Gy in 5 or 6 fractions, (2) single-lesion IMRT or VMAT plan created using multi-criteria optimization (MCO), (3) no prior radiation or no dosimetric impact from any prior radiation, and (4) 98% of GTV covered by >50 Gy. An A value of 0.9 was used to calculate the gEUD of normal liver (excluding GTV). The artificial neural network consisted of five layers including three hidden ones. It used six inputs: the x, y, and z coordinate of the center of PTV relative to the center of total liver (including GTV), the distance between the two centers, and the volume of the PTV and total liver. The model was tested on new consecutive patients since February 2018. Since the model is bounded to input ranges, a new patient with input outside the initial range cannot be used test the model. For this reason, five new patients were added to the learning data before ten qualified new patients were found to test the model. All training and test patients were treated on two different linear accelerators, both with 5-mm leaves in central 20×20 cm2 field. The model accurately predicted normal liver gEUD for five patients, with an error of <3.2% (<0.35 Gy). For the other five, the error ranged from 7.9% to 12.9% (0.76 to 1.33 Gy). On average, the error was -0.8±8.4% (-0.14±0.98 Gy). The error increased when any input approached the limit in the under-sampled boundary region. For the patient with the largest error, the x of PTV center was -7.23 cm, only 0.08 cm away from the lower limit. For the patient with the second largest error, the y of PTV center was 7.16 cm, only 0.75 cm away from the upper limit. The model will become more robust once it learns from those patients with inputs in the boundary regions. The calculation for a new patient takes <1 second on a regular PC. The model can be exported as a Python function and scripted in common contouring and planning systems. A machine learning model was built to predict normal liver gEUD in 50-Gy liver SBRT, only using the geometric properties of liver and PTV. The prediction was within 13% (1.3 Gy) of the actual value for ten consecutive prospective patients. The model can facilitate the Rx selection in liver SBRT, promoting efficiency for planners and clinicians.

Development and validation of a deep learning‐based model for predicting burnup nuclide density

Prediction of burn‐up nucleus density based on machine learning

A Machine-Learning Model Using Artificial Neural Network to Facilitate Liver SBRT Prescription Selection by Predicting Normal Liver Geud Based on Geometric Properties of Liver and PTV

Study of Assembly Nuclide Density Prediction Based on Data Mining Technology

Global prediction of nuclear charge density distributions using deep neural network

Prediction method for thermal-hydraulic parameters of nuclear reactor system based on deep learning algorithm

Application of the Deep Learning Algorithm to Identify the Spatial Distribution of Heavy Metals at Contaminated Sites

Predicting BWR Criticality with Data-Driven Machine Learning Model

Research on the Preliminary Prediction of Nuclear Core Design Based on Machine Learning

Machine Learning Modelling of Decay Heat Removal in High Temperature Gas-Cooled Reactor

High-risk nuclide screening and parameter sensitivity analysis based on numerical simulation and machine learning

Neutron transport calculation for the BEAVRS core based on the LSTM neural network

Development and Application of Data-Driven CHF Model in System Analysis Code

Prediction of boiler gas side effective heat transfer coefficients using mixture density networks and historic plant data

Diagnosis and Prediction for Loss of Coolant Accidents in Nuclear Power Plants Using Deep Learning Methods

Regression Transients Modeling of Solid Rocket Motor Burning Surfaces with Physics-guided Neural Network

Deep Learning for Nuclear Physics

A New Deep-Learning-based Model for Predicting 3D Radiotherapy Dose Distribution In Various Scenarios

Prediction Model for Aero-engine Combustor Temperature Field with Physical Constraints of High Temperature Deviation

Towards robust prediction of material properties for nuclear reactor design under scarce data -- a study in creep rupture property

Machine learning prediction of BLEVE loading with graph neural networks