Medical Image Processing Validation and Accuracy Prediction:from Clinical Exploitability to Brain Connectivity

da aggiungere con l versione definitiva della tesi

Validation and accuracy assessment are the main bottlenecks preventing the adoption of many medical image processing algorithms in the clinical practice. In the classical approach, a-posteriori analysis is performed based on some predefined objective metrics. In this thesis, a different approach is proposed. The basic idea consists in predicting the accuracy that will result from a given processing on a given type of data based on the identification and characterization of the sources of inaccuracy intervening along the whole chain. To this scope we have developed a complete validation framework, that exploits not only the classical validation and accuracy assessment metrics and approaches, but also methods arising from risk analysis and graph theory. It is proposed a proof of concept in both single-task and complex clinical scenarios. For what concerns simple clinical scenarios, we have focused our work on the validation of Magnetic Resonance (MR) images registration, highlighting not only the registration framework related aspects, but also considering the possible sources of inaccuracies that can rise from the quality of the images themselves. In particular MR images are normally corrupted by random noise from the measurement process complicating the automatic feature extraction and analysis of clinical data. The accurate registration of images observed in additive noise is a challenging task. The noise can increase the number of misregistered regions, and decrease the accuracy of subpixel registration. Besides, considering complex scenarios, we have investigated the problems related to a Diffusion MR images analysis pipeline, in order to detect the critical aspect that affect the accuracy of the results the most, especially in terms of robustness and reproducibility. In this context we have taken into account the whole processing system that comprises segmentation, registration and data reconstruction steps by stressing both each single module and the framework itself on the whole. Results show that the proposed model not only provides a good prediction performance, but also suggests the optimal processing approach in terms of accuracy, time load and robustness.