Brain diseases are the leading cause of physical and mental disability in modern society. Timely diagnosis and effective treatment of these diseases are possible only with an objective, highly-sensitive and reliable tools for observing the patients. These tools are also the basis for understanding the cause and progression of the disease as well as for evaluating the effectiveness of new drugs.
One of the neurodegenerative diseases that affects young adults and spans across their entire lifetime is multiple sclerosis (MS). The clinical symptoms of MS are difficult to assess in an accurate and reproducible manner due to their highly heterogeneous manifestation. On the other hand, one of the morphologic characteristics of MS is multi-focal nerve injury. The inflammatory response that accompanies the multi-focal neural injury is clearly visible in magnetic resonance images (MRI) of the brain, where multiple local inflammatory regions also known as lesions can be observed. For that reason, magnetic resonance imaging became a standard imaging technique for diagnosis and monitoring of MS.
Pathological brain processes can be objectively evaluated by defining the number, size, shape and anatomical position of brain structures, e.g. by manually outlining these structures. This task is very challenging, time-consuming, but most of all subjective and thus unreliable. On the other hand, automatic analysis of MRI can provide much faster, more accurate and reproducible results.
In recent years, several techniques for automatic segmentation of pathological structures have been proposed. Currently, the most successful methods are based on machine learning techniques, such as random forests and convolutional neural networks. We have quantitatively evaluated both methods and objectively evaluated and cross-compared their performance on an MRI image dataset of patients with MS. Best results were obtained with the convolutional neural networks. We also found that random forests are very sensitive to the pre-processing of MRI images, while this has almost no impact on convolutional neural networks. Furthermore, we found that the type of MRI scanner and the distribution of lesion sizes also have an important impact on the lesion segmentation.
Besides the pathological structures, a high-quality segmentation of healthy tissue is crucial to monitor the course and treatment efficacy of MS. From the literature it is known that with the presence of lesions on MRI images, the segmentation of healthy tissue is generally inadequate. For this purpose, we have developed a lesion filling approach, which allows the delineation of healthy tissue to be substantially improved. The method was objectively evaluated and compared to three state-of-the-art lesion filling methods. Experimental results revealed that the proposed method significantly outperformed the other three methods.
Automatic analysis of MRI images is not yet common in clinical practice. The introduction of such technology would clearly help the radiologists to work more efficiently and extract more information from the MRI images. The information in the form of quantitative measurements of healthy and pathological structures of the brain could offer important additional insight into disease activity.