Convolutional neural networks for lesion segmentation in brain magnetic resonance imagesMACERL, JURE (Avtor) Špiclin, Žiga (Mentor) multiple sclerosismagnetic resonance imagingmachine learningconvolutional neural networksBrain is one of the largest and most complex organs in the human body. It consists of many specialized and interconnected substructures that work together, and is responsible for crucial tasks like motor function, vision, speech, memory, cognition, etc. Therefore, any damage to the brain is reflected in physical and mental disabilities. In the context of this thesis we will focus on multiple sclerosis (MS), a progressive neurological disease that affects the central nervous system. The hallmark of the disease is scar brain tissue, also known as lesions, which are visible on magnetic resonance (MR) images. As a noninvasive diagnostic method, the MR imaging is an important paraclinical tool used to diagnose and to monitor MS disease activity. Early diagnosis and objective patient observation based on MR may estimate the disease progression and the effectiveness of immunomodulatory therapies. Disease diagnosis, progression and response to therapy can be objectively determined by measuring the volume, size, and location of the lesions in brain MR images. This may be done by a neuroradiologist by manually outlining the lesions. However, this task is time-consuming and subject to human error and huge bias. In recent years, state-of-the-art automated methods for segmenting the lesions in MR images have emerged and promise to provide faster and reproducible results. With increasing computational power and wide availability of machine learning (ML) frameworks, and their ability to learn and improve performance on specific tasks, the ML has become widely used to solve nowadays challenging engineering problems like autonomous-driving, speech recognition, image recognition, and segmentation. Convolutional neural networks (CNN), a type of ML algorithm most commonly applied to image recognition problems, have shown outstanding performance in various applications. The CNN is a supervised learning method, in which a mapping function from input to output is established based on labelled training samples. We are motivated to apply CNN in context of lesion segmentation since the CNNs have previously achieved superhuman performance level on certain classification and semantic segmentation tasks in the context of biomedical image analysis. In this thesis, we designed, tested and validated an ML based framework for automatic lesion segmentation. We used CNNs for semantic segmentation of lesions in MR images. In the first part of the thesis, we focused on segmentation of white matter lesions and in the second part on segmentation of cortical gray matter lesions. The framework allowed us to run experiments with several training data augmentation techniques, network architectures and in the context of various segmentation tasks, using transfer learning. The aim of the thesis was to test and validate the CNNs for quantification of MR images and to obtain benchmark performance metrics that determine whether such an approach can be used to objectively assess the disease status. First, we tested different augmentation techniques like intensity normalization, sagittal flip and random B-spline based spatial deformations. Next, we tried to find the best network architecture for the lesion segmentation task, with a hypothesis that a multi-pathway approach could address the trade-off between the amount of contextual low-resolution information needed to detect lesions and the amount of high-resolution information needed for accurately delineate lesions. Finally, we applied transfer learning to address another task, i.e. segmentation of cortical gray matter lesions. Objective validation of the CNNs for the mentioned lesion segmentation tasks was carried out on MR image datasets of 467 MS patients, who were imaged at 7 different sites on scanners from all three major vendors. We found that augmentation does not substantially increase segmentation performance, when compared to using multisite datasets, and that input image quality has a big impact on accuracy. The white matter lesion segmentation results show the benefit of our novel multi-pathway CNN architecture that achieved good performance across multiple scanner datasets, improving performance up to 13%. Finally, transfer learning as a means of migrating CNN weights from models trained for white matter lesion segmentation to models for gray matter lesion segmentation was found promising. We found that it is difficult to segment lesions in the gray matter due to their small size and rather poor contrast, but also due to a rather small number of available training and test cases. However, the results are promising and more data would be required for objective validation. In conclusion, the white matter segmentation results indicate that CNNs are robust against inter-scanner differences and may be applied for the quantification of lesions on standard clinical MR images used to diagnose and monitor MS patients.20182018-12-21 09:40:03Magistrsko delo/naloga105875VisID: 44901sl