Institutes outside Greece
Artificial intelligence applications in radiotherapy: the role of the FAIR data principles
Abstract
dc:descriptionIn recent years, research on artificial intelligence (AI) in the radiotherapy (RT) domain has begun to monopolise the interest of different clinicians and researchers.This will likely make routine RT clinical procedures more reliable and effective with a direct benefit to the patients. However, the clinical integration of these AI-based applications is still lacking as several steps still need to be included in the AI clinical implementation roadmap starting from the research stage and ending up to the clinical integration. One of these crucial steps is the adoption of the FAIR (Findable, Accessible, Interoperable, Reusable) data prin-ciples among the different data structure/archive systems of the radiotherapy centres. The FAIR principles can establish an interoperability framework for the reusability of RT multi-source data that can potentially decrypt valuable information for prognostic or diagnostic research and clinical purposes.This thesis focuses on the introduction of AI techniques and the FAIR data principles in four RT subdomains: i) medical imaging, ii) prediction modelling or RT related outcomes, iii) quality assurance (QA) of RT treatment planning and iv) the implementation of the FAIR data principles.Medical imagingIn the first part of the thesis we focused on the medical imaging domain. We investigated how RT can be transformed into a personalised treatment modality using imaging derived features that can decrypt valuable information that is not visible by the human eyes using machine learning (ML) techniques (ie. radiomics). Specifically, in chapter 2 we provided a roadmap for the clinical implementation of radiomics-based prediction models in the clin-ic, identifying the pitfalls and uncertainties encountered in the radiomics methodolo-gy/pipeline followed by the researchers. Furthermore, based on the pitfalls and uncertain-ties we presented, we proposed a standardisation framework with all the necessary tech-nical aspects that should be taken into account in the design/development of a radiomics study.One main take-away from chapter 2 is that the reproducibility of radiomics stud-ies/findings is one the main requirements for the standardisation and implementation of the radiomics concept in the clinic. Therefore, in chapter 3 we provided a publicly available dataset consisting of Computer Tomography (CT) phantoms (suitable for radiomics stud-ies) scans from three different Dutch RT centres, having as a goal to promote the repro-ducibility and interoperability of radiomics studies. Prediction modelling or RT related outcomesIn the second part of the thesis we focused on the prediction modelling of RT outcomes using AI algorithms. Taking into account the patients data privacy regulations, in chapter 5, we implemented a radiomics-based federated decentralised multicentre study, using tNon Small Cell Lung Cancer (NSCLC) patients. Having as a base the FAIR transformed clini-cal and radiomics features-based data, we validated a radiomics signature that predicts the 2-years overall survival of NSCLC patients without the exchange of patients data, com-paring the results of the centralised analysis. The study showed that the performance of the radiomics signature of the decentralised approach was not significantly different than the centralised one. Furthermore, with this study, the significance of the FAIR transfor-mation of data for federated learning radiomics studies is underlined, implementing a privacy preserving infrastructure without the data exchange of patients.In chapter 6, we performed an independent validation of the logistic regression-based Normal Tissue Complication Probability (NTCP) model that predicts the six months ≥ 2nd grade dysphagia for head and neck cancer patients. This model is part of the Dutch Na-tional Indication Protocol for Proton therapy (NIPP) for the selection of patients candidates for proton therapy (PT). In this study, we showed that logistic regression models need a specific validation approach in independent cohorts, examining the potential update of the different components of the logistic regression models (intercept, slope and predictor coefficients) through the AI-based approach of the closed testing procedure (CTP). The CTP in combination with the graphical assessment of the calibration curves of the different updated models, indicated that the original dysphagia NIPP model needed an update and a new updated calibrated model was selected. However, it is important to perform a fed-erated privacy preserving multicentre study using FAIR transformed datasets in the dif-ferent Dutch PT centres that can robust the results of the CTP.Quality assuranceThe RT treatment planning is a complicated procedure that requires the coordinated ef-forts of clinical and technical RT professionals. For the third part of the thesis, In chapter 7 we focused on the quality assurance (QA) of the RT treatment planning procedure by ex-ternally validating an AI-based method developed in the United States (US) using a Dutch independent patients cohort. This AI method using Bayesian Networks (BNs) has as a goal to early detect errors encountered in the verification phase of RT treatment planning and alert humans for possible erroneous variables included on it. The external validation using an independent Dutch patients cohort was not successful, due to the different technical characteristics of the treatment machines and software used in the different RT centres possibly. According to this study, further steps are required for the generalisability and reusability of AI-based systems focusing on the automatic error detection among different centres such as data preprocessing and the inclusion of more variables included in the treatment planning such as imaging-based data.Implementation of FAIR principlesIn chapter 4, expanding the work of chapter 3, we provided a set of four publicly available datasets that were used in a breakthrough publication in the radiomics community in 2014 in a machine-readable format. Specifically, using radiation oncology related ontolo-gies and semantic web technologies, we transformed multisource (clinical, radiomics fea-ture-based, and imaging) data in a FAIR format for enabling the automation of data pro-cessing by the machines with minimal human intervention.For the last part of the thesis, we focused on the implementation of the FAIR principles in the RT domain. In chapter 8, having as a goal to introduce the FAIR concept in the RT community, we provided an overview of the action points required from the different RT stakeholders for the introduction and adoption of the FAIR principles in the different data archive systems of the hospitals. Some of the advantages that a FAIR data transformation offers are the flexibility to adapt databases and the ability of machines to “read” the dif-ferent data for automated AI-based studies. In chapters 9-10, we provided FAIR data models structured in knowledge graphs using the data elements used for the Dutch na-tional registry of patients candidates for PT. Using publicly available ontologies and se-mantic web technologies we underline the significance of the flexibility and interoperabil-ity of the FAIR format using routine clinical data.
Degree
thesis:*- Grantor dc:publisher
- Institutes outside Greece
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kalendralis, Petros
- Καλένδραλης, Πέτρος
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53992