Three-dimensional dose prediction method and system for radiotherapy

Abstract

The invention provides a three-dimensional dose prediction method for radiotherapy. The method includes the following steps of A, collecting radiotherapy planning data of past cases, and preprocessingthe data to obtain a feature image capable of being trained by a convolutional neural network; B, calculating the minimum distance from each voxel in a planning region to a target region boundary toobtain a target distance map; C, establishing a three-dimensional dose prediction network based on the convolutional neural network; D, training the three-dimensional dose prediction network through the feature image in step A to obtain an optimal three-dimensional dose prediction model through cross validation; E, inputting data of a patient to be irradiated into the three-dimensional dose prediction model to obtain a three-dimensional dose distribution map. The invention also provides a prediction system based on the method. The three-dimensional dose prediction method and system for radiotherapy have the advantages that by introducing deep learning of the convolutional neural network, the dependence on the personal experience of physicists is reduced, and the error of manual predictionis reduced.

Description

technical field [0001] The invention relates to the technical field of disease radiation quality planning, in particular to a three-dimensional dose prediction method and system for radiation therapy. Background technique [0002] Radiation therapy is one of the main treatments for cancer. With the development of advanced treatment methods such as IMRT (Intensity-Modulated Radiation Therapy, Intensity-Modulated Radiation Therapy) and VMAT (Volumetric-Modulated Arc Therapy, volume intensity-modulated), the quality of Linchuan plans has been significantly improved. However, making a high-quality clinical plan usually requires a lot of manual intervention, and physicists iteratively optimize the planning parameters based on their previous experience and feedback from doctors until the plan is clinically accepted, which leads to an increase in planning time. Studies have shown that the dosimetric features of high-quality plans are often strongly linked to their geometric anatom...

AI Technical Summary

Problems solved by technology

[0003] Current dose prediction work mainly focuses on predicting the dose volume histogram (DVH) of patients, which requires a lot of time to analyze the association between dose distribution and patient characteristics

In particular, it is necessary to extract manual features, such as spatial information of organs at risk (OAR), shape of target volume (PTV), distance histogram (DTH), overlapping volume histogram (OVH), etc., for dose prediction. This method is complex Sexual and very time consuming

At the same time, there may be other unexplored features that can significantly affect the prediction accuracy, which are seriously missing in the above methods

In addition, the prediction of patient DVH can only provide the statistical information of the dose distribution, but cannot provide the prediction of the detailed three-dimensional dose distribution of the patient.

Images