Studies on Machine Learning Aproaches to Proton Radiography

Abstract

AB S TRACT newlineProton therapy is one of the modern methods for cancer treatment newlinewith increasing demand globally. However, the problem newlineof range uncertainty which reduces the quality of proton treatments newlineremains to be unresolved. In this context, proton imaging newlineis a promising method in reducing the range uncertainty that newlinearises from the calibration curve that converts x-ray Hounsfield newlineunits (HU) to relative stopping power (RSP). Proton energyresolved newlinedose imaging (pERDI) is a recent technique in proton newlineimaging which is based on the energy dependence of the dose newlinedistribution delivered by a scanned proton beam. The main advantage newlineof this technique is the use of a single detector to measure newlinethe exit dose with multiple beam energies, which reduces newlinethe complexity of the setup. newlineThis research work aims to propose new approaches in the newlinefield of proton radiography using the pERDI technique. This newlinethesis focuses on the pERDI technique to generate proton radiographs newlineusing various approaches like the and#967;2-minimisation technique, newlinean analytical expression, and different machine learning newline(ML) techniques like neural networks from the Keras library, newlineExtreme Gradient Boosting (XGboost) algorithm and advanced newlinedeep learning (DL) technique. newlineThe first part of the thesis includes the studies on the achievable newlineaccuracy in water equivalent path length (WEPL) and RSP newlineusing the pERDI technique in homogeneous and heterogeneous newlinemedia. The results are promising, where the achievable WEPL newlineaccuracy in a homogeneous medium was within 1 mm using newlinethe Lynx detector, and the RSP is measured with a precision newlinebetter than 1.5% for all homogeneous inserts of tissue surrogates. newlineThe first results demonstrated that the WEPL and RSP newlinecan be estimated with clinically acceptable accuracy using the newlinevi newlinepERDI technique. The WEPL map generated from the anthropomorphic newlinehead phantom is compared with the xCT image using newlinegamma index analysis. The gamma index test failed in the newlinehigh heterogeneous regions indicating that the WEPL accuracy newlineis affected by the range-mixing effects.