|Title:||Assessment of treatment plan complexity using neural networks||Language:||English||Authors:||Schinerl, Judith||Qualification level:||Diploma||Keywords:||Neuronale Netzwerke; patientenspezifische Qualitätssicherung; Fluenzverteilungen; Strahlentherapie
Neural networks; patient specific quality assurance; Fluence maps; Radiation therapy
|Advisor:||Georg, Dietmar||Issue Date:||2019||Number of Pages:||77||Qualification level:||Diploma||Abstract:||
Machine learning and especially neural networks receive more and more attention in current scientific applications, as shown by the increasing number of publications. Especially in medicine and biomedical engineering, where human errors still prove to be a cause of failure, machine learning algorithms are used to overcome the limits of human decision-making. One field that may benefit from the recent developments is radiation therapy, as machine learning algorithms perform well on classifying image data. The standard clinical workflow at the Medical University of Vienna / General Hospital of Vienna (AKH) includes a quality assurance (QA) measurement in advance of each high precision patient treatment. Each treatment is planned beforehand using the software Monaco (Elekta AB, Stockholm, Sweden), which determines beam energies and the positions of collimators of the medical linear accelerator (linac) according to a desired dose distribution covering delineated regions of interest on computer tomography images. The QA measurement on the linac is then performed with a verification phantom, and finally planned and measured dose distributions are compared in order to ensure safe dose deposition in the patient. The gamma passing rate (GPR) serves as a measure of conformity of these two dose distributions. The GPR depends on the size, shape and location of the tumour in the patient and has to exceed a certain value in order for the plan to be regarded as safe to irradiate on a patient. In this thesis, the setup and training of a convolutional neural network (CNN) with the aim of classifying treatment plan data by estimating the GPR is described. Achieving this task with sufficient accuracy will enable a higher efficiency of the QA procedures and more efficient use of the medical linear accelerator. A neural network is a deep learning concept mimicking the information processing in human neurons by its layered structure. Each layer is composed of a number of nodes, called neurons. They are connected to other neurons in the previous and/ or subsequent layers, each connection being weighted by a weighting function. The training process involves the passing of input data with known output to the network, i.e. data that has been labelled according to their corresponding class. To obtain input images, 600 volumetric modulated arc therapy (VMAT) treatment plans for either prostate, gynaecological or head-and-neck (HN) cancer generated for either Versa HD or Synergy linacs (Elekta, Sweden) were extracted and assigned to one of three labels according to their GPR value. The planning data, i.e. the dose and positional information of all collimators contained in the planning data in Dicom (digital imaging and communication in medicine) format was then transformed into grayscale fluence maps, depicting the transmission of dose through the beam window. The complete dataset was separated into three smaller datasets designated for training, testing or evaluation purposes. During training, which was performed in Python using the framework TensorFlow, two datasets were used to set the weights in order to output the known label of each input accordingly following an optimisation operation. Several different models were trained, varying the learning rate, batch size and depth of network in order to improve achieved training accuracies and further testing the robustness of the resulting layer structure by switching and shuffling the datasets. The achieved training accuracies range from 57% to 69%, showing a large variation upon changing the layer sequence and parameters. Furthermore, robustness testing revealed large variations of accuracy upon switching the used datasets, leading to accuracies between 59% and 69%. Evaluating the best performing convolutional neural network on unknown data, i.e. the third dataset not used during training, resulted in an evaluation accuracy of 59.5%, showing a reduction of 10% compared to the training accuracy of 69%. Similar values can be found in recent literature evaluating fluence maps of radiotherapy treatments according to the associated GPR1. Since the obtained results only offer a first insight on the performance and behaviour of CNN, various approaches to increase the achieved accuracies and enhance network robustness have been identified. Improvements with respect to accuracy and robustness are necessary for utilizing these CNNs in a clinical workflow but go beyond scope of this work, as the objective was to identify general mechanisms and problems of neural networks in radiotherapy. The results obtained in this thesis show the potential of CNNs acting as a promising new approach for applications in quality assurance in radiation therapy.
|Library ID:||AC15325060||Organisation:||E141 - Atominstitut||Publication Type:||Thesis
|Appears in Collections:||Thesis|
Show full item record
Files in this item:
|Assessment of treatment plan complexity using neural networks.pdf||3.83 MB||Adobe PDF|
checked on Feb 18, 2021
checked on Feb 18, 2021
Items in reposiTUm are protected by copyright, with all rights reserved, unless otherwise indicated.