[Proposal-ML] Segment Vasculature in 3D Scans of Human Kidney
Keywords: machine learning, 3D image segmentation, blood vessel segmentation, denoising pertaining, U-Net, biomedical imaging, fine-tuning
TL;DR: The project focuses on improving blood vessel segmentation in high-resolution HiP-CT kidney images by using a U-Net3D model with denoising pretraining and fine-tuning.
Abstract: The Vasculature Common Coordinate Framework (VCCF) seeks to create a detailed map of the human vascular system to enhance understanding of cellular interactions. Progress is hindered by labor-intensive annotation processes, with current methodologies taking up to six months per dataset. Currently, machine learning models struggle with generalization due to anatomical variability and varying image quality from Hierarchical Phase-Contrast Tomography (HiP-CT), which produces high-resolution 3D imaging essential for accurate vascular mapping. This project aims to develop a novel machine learning model for the segmentation of blood vessels in HiP-CT images of human kidneys, building on a successful Kaggle competition. We select a straightforward U-Net3D model from the second-place solution for its effectiveness. To improve performance given the sparsely labeled dataset, we apply denoising pretraining to the decoder and then use a densely annotated dataset to fine-tune. The optimized models will be evaluated through Kaggle’s Late Submission, with the best-performing version selected for final reporting. This research aims to significantly improve vascular mapping within the VCCF initiative.
Submission Number: 49
Loading