PDF, English - main document Download (23MB) | Lizenz: Proton Dose Calculation using Artificial Neural Networks by Neishabouri, Ahmad underlies the terms of Creative Commons Attribution-NonCommercial-NoDerivatives 4.0

Abstract

This doctoral research investigated the application of Artificial Neural Networks (ANNs) and deep learning techniques for estimating proton dose distributions in particle therapy. The study focused on evaluating the suitability of ANNs as a dose estimation method. The primary objectives were: 1) To develop and evaluate ANN models, particularly Long Short-Term Memory (LSTM) networks, capable of capturing the spatial dependence and heterogeneity of patient anatomy in mapping CT images to dose distributions. 2) To quantify dose prediction uncertainties using Bayesian LSTM models. 3) To investigate the ability of ANNs to meet the goal of real-time adaptive proton therapy by creating a dose estimation engine and models.

The initial feasibility study demonstrated that LSTM networks could effectively learn the supervised task of proton dose estimation, correlating spatio-temporal features between input CT images and simulated ground truth Monte Carlo dose distributions. It was shown, that the LSTM networks outperformed other variants of Recurrent Neural Networks-based models in terms of accuracy and computational efficiency and exhibited generalization capabilities for different energies and patient anatomies. The millisecond run-time per pencil beam suggested the potential of these models to generate full-field dose distributions for real-time adaptive proton therapy.

To address the frequently voiced concern of 'explainability' and to quantify model prediction accuracy for clinical translation, the Bayesian LSTM model, BayesDose, was developed. By incorporating probabilistic elements, BayesDose enabled the assessment of prediction uncertainties, providing confidence intervals and identifying potential sources of error.

Based on the lessons learned from the feasibility study, a custom dataset was created to investigate the deep learning models in a real patient scenario, with previously delivered treatment plans in the head region exhibiting a full range of heterogeneities, from air cavities to dense bony structures. Due to the challenges posed by the new dataset and the wide proton beam in HIT, the initial LSTM models were unable to correctly learn the spatio-temporal features. To address these challenges, two custom-designed, physics-informed variations of LSTM were proposed: LSTM133 and CC-LSTM. LSTM133 facilitated the training of higher dimensional inputs and outputs, while CC-LSTM was designed from scratch, tailored specifically to the characteristics of proton interactions, significantly improving accuracy and computational efficiency. CC-LSTM processed the input in three steps: 1) dimensionality reduction of the input using strided convolutions, followed by spatial feature extraction using two Convolutional Neural Network (CNN) layers 2) spatial feature fed to a ConvLSTM cell, that by updating its cell state and hidden state, propagates the spatial features temporally, and 3) dose distribution prediction using a three layer CNN backend. This model achieved the computational speed necessary for real-time adaptive proton therapy (APT) and outperformed the current state-of-the-art in terms of accuracy and run times, setting a new benchmark for deep learning-based particle therapy dose calculation. Moreover, this thesis involved a comprehensive development aspect that allowed forward calculation of an entire field, end-to-end, via only the treatment plan, and the corresponding patient CT acquisition. This forward calculation operate in four steps, 1) The RSP cubes are extracted from the CT based on the treatment plan 2) the BEV dose distributions of all pencil beams are inferred in one feed forward of the model 3) these dose distribution are then accumulated based on the plan, on the GPU, in the BEV, and 4) The dose distribution is back-interpolated to the CT grid with the desired resolution. The design and evaluation of methods, models, workflows, and code libraries generated in this study have facilitated fast dose calculations for APT. The proposed dose calculation engine could carry out dose distribution of iso-energy surfaces in \emph{under one second} on a desktop computational system, with highly conformal accuracy comparing to the ground truth MC simulations. Such performance ensures that dose estimations could be performed rapidly enough to adapt to dynamic changes in patient anatomy during treatment, given the availability of in-room monitoring systems such as cone beam CT, vision RT, MRI, and read from the available in-room imaging systems. This way, the proposed engine can evaluate updated dose distributions based on discrepancies in dose delivery read from machine's guiding systems such as beam application and monitoring system or based on intra- and inter-fractional changes in the patient anatomy.