Neurological diseases are among the most common causes of mental and physical disabilities in the modern society and as such present a huge socio-economic impact. At the onset and during the progression of these diseases, for instance in Multiple Sclerosis, the underlying pathological processes often result in the occurence of scar tissue or lesions in the white matter. These lesions are best observed by Magnetic Resonance imaging. Quantification of white matter lesion volume and count from the MR images therefore represents an important biomarker for diagnosis, prognosis and treatment followup of neurological diseases. So far many automated methods for the MR image quantification have been developed. However, an objective an extensive validation of such methods on real, clinical image dataset is a challenging task since it requires accurate and reliable reference delineations of white matter lesions, which can only be acquired by manual delineation of each lesion. To facilitate the creation of such reference delineations we proposed a protocol, in which multiple raters created a consensus based reference delineations of white matter lesions. Validation of the proposed protocol shows that it results in a more accurate and reliable reference delineations of lesions in comparison with the delineations made by any single rater. The protocol was used to create reference delineations of white matter lesions for a database of 30 MS patients, which consisted of T1- and T2-weighted and FLAIR images MR images acquired on a 3T MR machine. The same protocol was also employed to build a second, longitudinal database of 20 MS patients with two MR studies per patient and corresponding reference white matter lesion changes delineations. Each study consisted of T1- and T2-weighted and FLAIR images MR images acquired on 1,5T Philips MR machine.
Using the longitudinal MR database with reference lesion change delineations we validated and
compared state-of-the-art methods for lesion change detection based on longitudinal analysis
of intensity variations. Validation was performed using standard metrics found in the literature
with the addition of a new metric – regional Dice coecient, which allowed the analysis
of methods’ performance with respect to the size of each particular lesion change. Obtained
results were not nearly as good as the ones reported by the original authors, which suggests
that the performances of the evaluated change detection methods might either be very dependent on the MR image acquisition or dependent on the accuracy of ground truth segmentation.
In order to facilitate validation of existing and newly developed change detection methods the
aforementioned datasets were publicly disseminated.
Machine learning algorithms can be used to determine the set of optimal features required for
white matter lesion change detection. Identification of such features helps us to better understand, which imaging information is the most important for change detection. Using a random forest classifier we evaluated and compared the importance of various features for lesion change detection, assessed the reliability of the estimated feature importance and, finally, determined the ability to generalize feature importance estimation to other classifiers. The selection of optimal features may influence the setup of MR acquisition parameters and sequences needed for lesion change detection and as such can be used to optimize MR acquisition, e.g. to remove MR sequences unnecessary for change detection. Such optimizations could potentially decrease the cost and time required to acquire an MR patient study, which would benefit both the patients and the medical personnel. The proposed feature selection method can also be used to develop change detection methods with higher accuracy and specificity, which, in turn, would enable clinicians a better insight into disease progression, increase their understanding of the underlying pathology and allow them to make timely and fully-informed decisision in case of ineective treatments.
|
