Proton Thermoacoustic Signal Traveltime Extraction Method Based On Deep Learning

Radiotherapy is one of the most important treatment methods for malignant tumors.Traditional radiotherapy is mainly based on photons and other basic particles.The high energy produced by particles when they are injected into human body will effectively kill tumor cells.However,the depth dose distribution of traditional radiotherapy with photons in human body decreases exponentially with the depth of incidence.This will make the normal organs and tissues in the photon beam pathway suffer from different degrees of radiation damage,and even lead to secondary cancer.How to ensure the effective killing of tumor cells and reduce the damage to normal organs and tissues as much as possible is a problem that scholars in the field of radiotherapy have to consider.However,radiotherapy with protons or heavy ions as the basic particles brings the possibility to solve this problem.When the proton beam or heavy ion beam inject into human body,its depth dose distribution is Bragg curve,and there is a fixed Bragg peak.Before the Bragg peak,the dose of proton beam deposition is less,and after the Bragg peak,the dose of proton beam deposition is close to zero.As long as the Bragg peak of the proton beam can be accurately located in the tumor target area,the tumor cells can be killed while the damage to normal human organs and tissues can be reduced.At present,the most widely used Bragg peak localization techniques are positron emission tomography(PET)and prompt gamma.However,it is expensive to build such treatment environment with low economic benefit ratio,and it is usually difficult to realize on-line real-time monitoring of Bragg peak.The technique of locating the Bragg peak using a proton beam thermoacoustic signal can theoretically monitor the Bragg peak online in real time,and the cost of setting up a treatment environment is low.In the process of dose deposition,proton beam can induce columnar α wave(from the deposition area before Bragg peak)and spherical γ wave(from the deposition area of Bragg peak).The travel time of γ wave is an important parameter of Bragg peak location technology using proton beam thermoacoustic signal.The accuracy of the travel time of γ wave is related to the accuracy of Bragg peak location.However,the traditional method of extracting travel time of γ wave usually has a "blind area" which is difficult to extract travel time,and the level of travel time error of γ wave extracted in different directions of Bragg peak is inconsistent,which has obvious angle dependence.In this paper,a method based on deep learning is proposed to extract the travel time of γ wave from the thermoacoustic signal of proton beam.Based on U-net,a U-type neural network model is designed in this paper to recognize and process the proton beam thermoacoustic signal.In addition,k-Wave MATLAB toolbox is used to simulate the propagation process of proton beam in the non-uniform acoustic velocity model,and the training data set containing 271,700 original proton beam thermoacoustic signal samples is collected.After adding noise to the training data set and labeling with the travel time of γ wave,it is put into the U-type neural network model for learning and training.After the training of the U-type neural network model,three groups of experiments are set up to test the γ wave travel time prediction performance of the U-type neural network mode,including the measurement of the travel time extraction of the proton beam thermoacoustic signal with different noise levels,the angle dependence test of the travel time error of the γ wave and the generalization ability test of the U-type neural network.The experiments’ results show that the U-type neural network model in this paper is comprehensive,accurate and has good anti-noise ability.At the same time,there is no obvious angle dependence of the travel time error of the γ wave extracted by the neural network model,so there is no need for further compensation.In addition,the neural network model also has a certain degree of generalization ability.For the proton beam thermoacoustic signals from different acoustic velocity models,it can also accurately and comprehensively extract the travel time of γ wave.