This master's thesis addresses the issue of automatic morphometry of the fetal head based on magnetic resonance (MR) images. The main objective of the research is to improve prenatal diagnosis of ventriculomegaly (VM), which is one of the most common neurological anomalies in fetuses. VM is characterized by increased lateral brain ventricles with diameter of 10 mm or more and is often associated with potential neurological problems after birth. Although VM is frequently diagnosed through ultrasound examination, MR images are more precise in detecting subtle abnormalities and are therefore essential in further assessments.
The thesis focuses on the use of advanced deep learning methods to automate the process of measuring fetal biometric features from MR images. In clinical practice, these measurements are performed manually by radiologists, which is a time-consuming and sensitive process, prone to errors due to subjectivity or poor image quality. To improve diagnostic accuracy and reliability, an automated measurement process was proposed.
In the first part of the research, methods for super-resolution reconstruction of 3D MR images from 2D MR stacks were analyzed. This step is crucial as it enables precise visualization of brain structures for further processing. The study used a dataset from the University Medical Centre Ljubljana, which included 2D MR image stacks of both healthy cases and cases with different degrees of VM. The gestational age of the fetuses included in the study ranged from 22 to 36 weeks. For reconstruction, the NiftyMIC algorithm was employed. In the second part, a deep learning algorithm was applied to the reconstructed 3D MR image of the fetal brain to localize the anatomical landmarks needed to acquire 11 standard biometric measurements. The algorithm was trained on 81 images obtained from the first step. In the third part, we evaluated the ability to differentiate between healthy and pathological measurements. As input, we used the measurements obtained in the second step and calculated percentiles according to guidelines for healthy fetuses at different gestational ages. Decision trees were then used to analyze the diagnostic success in distinguishing between different forms of VM and healthy cases. Additionally, the algorithm was capable of classifying VM based on its severity, dividing the ventricles into three subgroups: mild VM (10-12 mm), moderate VM (13-15 mm), and severe VM (above 15 mm).
We found that the success of the 3D MR image reconstruction depends on the accuracy of the image acquisition, particularly in terms of the orthogonal positioning of the fetus and the presence of artifacts in the 2D images. Successful reconstruction was achieved in 73.63% of cases. Using the proposed method for detecting 22 anatomical landmarks on the reconstructed 3D MR images, we achieved an average localization error of 3.73 ± 1.80 mm. This result was slightly higher than in similar studies that focused on specific points rather than a comprehensive analysis. When determining the diagnosis using the decision tree, we found that the distinction between healthy and pathological cases was very high, though the accuracy decreased, owing also to limited number of cases, as the diagnosis task was made more complex by considering the VM subdiagnoses.
With this thesis, we aimed to demonstrate the advantages of using deep learning methods on clinical cases as a complementary tool to classical diagnostics. Our goal was to provide not only diagnostic results but also the underlying information on which the diagnosis was based, allowing for a quicker insight into fetal development.
|
