Introduction and Problem Motivation: Magnetic resonance (MR) imaging enables non-invasive measurements of anatomical structures in fetuses and monitoring its growth and development, thereby enabling early diagnosis of pathological conditions. For instance, ventriculomegaly refers to the enlargement of the lateral ventricles, which is diagnosed when the sagittal diameter of at least one lateral ventricle is 10 mm or more. Based on the width of the ventricles, we distinguish between mild, moderate, and severe forms of ventriculomegaly, with larger ventricular diameter associated with worse postnatal prognosis. The diagnosis is usually made based on manual linear measurements. In this master's thesis, we aim to establish automated volumetric measurements as effective for diagnostic purposes.
Data: The master's thesis addresses two databases: UKC and Kispi. The UKC database contains 2D MR image stacks of healthy cases and cases of various forms of ventriculomegaly. The gestational age in the database ranges from 22 to 36 weeks, with an average age of 30 weeks. The Kispi database was created as part of the FeTA challenge, which focuses on fetal brain segmentation. The Kispi database includes 80 3D MR images in a standard anatomical position, most of which are pathological cases, though specific pathology is not specified. The gestational age here ranges from 20.1 to 34.8 weeks, with an average age of 26.9 weeks. For each case, we also obtained manual segmentations of 7 brain tissues.
Methods: We developed a multi-step computational pipeline that includes super-resolution reconstruction of 3D MR images in a standard anatomical position, automatic brain segmentation, volumetric measurements, analysis of results, and models for diagnosis of ventriculomegaly. We performed the reconstruction of 3D MR images in a standard anatomical position based on 2D image stacks from the UKC database using the NiftyMIC algorithm. For the segmentation of both databases, we used the CNN algorithm dubbed BOUNTI. Volumetric measurements were obtained by counting labeled voxels. We checked the uniformity of centiles on healthy fetuses, which were obtained with respect to independent external reference database, with the Kolmogorov-Smirnov test. We analyzed the course of volumetric measurements in relation to gestational age. In the diagnosis, we aimed to test the ability to differentiate measurements between healthy and pathological cases, using AUC values (for individual measurements) and UMAP projection (for groups of measurements). We analyzed the diagnostic performance using easily interpretable decision trees, reflecting the choice of measurements and associated threshold values.
Results: The success of 3D MR image reconstruction depended on the presence of artifacts in the 2D images. Segmentation was poorest in deep gray matter due to different definitions or assumptions between the BOUNTI algorithm and reference segmentations, and in extremely enlarged lateral ventricles. A uniform distribution of measurement centiles on healthy cases was achieved in the UKC database for the entire volume of lateral ventricles and the left ventricle, while in the Kispi database, it was achieved for white matter, the left lateral ventricle, and the cerebellum. Volumetric measurements of the left, right, and total volume of lateral ventricles deviated from healthy cases. Lateral ventricles achieved the highest differentiation capability between healthy and pathological cases in both databases. UMAP projection demonstrated the good differentiation capability of the uniform feature centiles. The UKC database projection effectively differentiated between healthy cases and various forms of ventriculomegaly, but could not distinguish between unilateral, asymmetric, and symmetric forms. The data from the Kispi database clustered into several groups, but due to the lack of specific diagnoses, we could not assess the effectiveness of the projection. UMAP also reveals the possible impact of super-resolution image reconstruction algorithms. Diagnosis with the decision tree effectively differentiated between healthy and pathological cases in both databases, but the classification performance decreased with more detailed case labeling.
Conclusion: Differentiation of volumetric fetal measurement was very good at distinguishing between healthy and pathological cases. The use of uniform feature centiles yielded meaningful results, with improved differentiation between severe versus moderate and mild forms of ventriculomegaly. However, categorizing pathologies based on whether ventriculomegaly is unilateral, symmetric, or asymmetric achieved poor results, mainly due to the lack of cases in specific categories.
|
