ERROR ESTIMATION IN QUANTITATIVE MEDICAL IMAGE ANALYSISMADAN, HENNADII (Avtor) Pernuš, Franjo (Mentor) -An impressive improvement of the effectiveness of medical care evident in the recent decades is to a large extent driven by the progress made in the fields of medical imaging and analysis. A hallmark characteristic of this trend is the transition from a purely visual, qualitative assessment of the medical images to a computational and more quantitative assessment, which involves in vivo image-based measurements. In the domains of disease diagnosis and monitoring, treatment efficacy assessment, but also in surgery and radiotherapy planning and execution, the clinical workflow is becoming increasingly dependent on image-derived measurements (i.e. imaging biomarkers). Some of the quantitative imaging biomarkers have already become well established as surrogates of clinical outcomes. The values of these imaging biomarkes may directly impact the decision-making process | hence, the accuracy and precision of the methods that extract the measurements from the images need to be rigorously validated. Problems of objective validation and comparison of measurement methods feature prominently in the medical imaging discourse. In registration and segmentation | the two major fields of image analysis, the state of the art of method validation and comparison is based on reference measurements usually requiring some human involvement. In case of registration it is the detection and manual localization of fiducial markers. For segmentation it is the manual delineation of anatomical structures by expert radiologists. Certain problems are inherent in this approach: humans are subjective { measurements by different experts usually disagree, they are error prone | they get distracted and tired, and their time is costly. When human errors in validation standards propagate to medical practice they acquire a potential to cause costly damages. The patients, medical care establishments and the economy at large are all impacted by the consequences of these errors. Strategies to predicting and preventing the measurement errors and cutting the costs associated with validation and comparison of measurement methods are discussed in this Thesis. A direct strategy to alleviate the costs of the burdensome manual reference creation is through automation. Such strategy was applied in the first contribution of this Thesis using a novel automated computational approach to gold standard reference dataset creation for validating rigid-body registration of pre-operative 3D and intra-operative 2D images. Therein, the use of automatic image analysis pipeline eliminated the need for human interaction and manual input, previously required in a semi-automated approach. This has significantly improved the registration accuracy as validated on intra-operatively acquired 3D and 2D images of twenty patients with cerebral aneurysms and arteriovenous malformations. A different, more inventive, strategy is to validate the measurement methods without ever creating a reference, through advanced statistical inference. Two new reference-free Bayesian frameworks for estimating the systematic and random errors of an ensemble of (automated) measurement methods, are developed in this Thesis. They facilitate the validation and comparison of measurement methods without requiring costly reference measurements. A clear advantage of this strategy is that it eliminates the need for the reference measurements altogether and therefore annihilates the associated costs. For instance, in the image analysis domain, applying several automated methods to a certain dataset requires only computational resources, which is much cheaper than engaging an expert to manually create the reference. The two proposed frameworks were successfully validated on several synthetic and on relevant clinical datasets, involving imaging biomarkers of neurological diseases. Theoretical developments of one of the proposed frameworks allow to use it for advanced applications of estimation of latent true values of an unobserved quantity and selection of best predictors for it from a set of related biomarkers. In conclusion, the contributions of this Thesis do not only solve the practical problems of reference creation, but address the conceptual problems associated with reference based error estimation. The two proposed and validated Bayesian frameworks represent important theoretical advances in the emerging field of reference-free error estimation, making this methodology practical for measurement method validation, comparison and further beyond | for selection of best predictors of unobservable quantities and the their estimation.20182018-07-19 15:00:02Doktorsko delo/naloga102002VisID: 43021COBISS_ID: 12105812sl