CEREBELLUM ISOLATION WITH MULTI-MODALITY MRI IMAGES USING 3D Unet

Abstract

The cerebellum plays a crucial role in human life. It is involved not only in motor control but also in various cognitive functions. Understanding how the cerebellum grows and develops throughout the human lifespan helps researchers better comprehend its role in the nervous system and aids in diagnosing related diseases. The first step in analyzing the cerebellum is to isolate it from whole-brain images. This step is critical because any errors at this stage can affect downstream analysis, such as volumetric measurements or morphological comparisons. Over the years, researchers have developed various tools for this task, but current methods face two key limitations. First, most tools are based on templates derived from healthy adults, which are not suitable for other age groups, particularly neonates. Second, many existing tools rely on a single MRI modality (typically T1-weighted images), which limits their ability to accurately distinguish boundaries, especially in regions where contrasts are subtle or noisy. In contrast, multi-modal MRI data (e.g., T1-weighted and T2-weighted images) provide complementary information that can enhance tissue differentiation and segmentation accuracy.

This project aims to develop a robust deep learning model capable of distinguishing the cerebellum from surrounding brain structures. We employ a 3D U-Net architecture, which has proven effective in extracting features from medical images. The U-Net model is designed to capture both high-level semantic features and low-level spatial details, making it ideal for distinguishing the cerebellum from nearby brain structures. Our model takes as input two channels of cropped cerebellar regions from multi-modal whole-brain MRI scans.

We evaluate model performance using Dice Score and Hausdorff Distance. Compared to existing tools, our model achieves higher accuracy. In particular, the lower Hausdorff Distance indicates a reduced likelihood of mislabeling surrounding brain structures as cerebellum. Moreover, our model shows its capacity to process images from different age groups.