Segmentação de vasos sanguíneos com poucos dados via transferência de informação de forma

Abstract

Blood vessel segmentation is an essential task in various areas of biomedicine, such as ophthalmology and neurology. However, this task poses a significant challenge due to the scarcity of large annotated datasets, which consequently makes the manual annotation of new samples for training efficient deep learning models a costly endeavor. Transfer learning techniques have shown promise in overcoming these limitations, allowing the reuse of representations learned in one domain to improve performance in another, even when few examples are available in the target domain. The central hypothesis indicates that leveraging a shape prior of vessel-like forms, such as their tubular and branching characteristics, can lead to more robust and data-efficient models. In this context, this dissertation investigates the transfer of shape representations from a synthetic domain to the segmentation of medical images in a few-shot regime. To this end, we introduce VessShape, a methodology for generating a large-scale synthetic dataset designed to instill a strong shape bias in segmentation models. VessShape images combine procedurally generated tubular geometries with a wide variety of textures, encouraging models to learn shape cues over appearance features. The models pre-trained with VessShape were then fine-tuned and evaluated on two real-world datasets from different domains. The results demonstrate that the approach achieves strong segmentation performance in few-shot scenarios, requiring only a small number of samples for fine-tuning. Additionally, the models demonstrate a significant zero-shot learning capability, proving able to segment vessels in unseen domains without any target-specific training. These results support that pre-training with a strong shape bias constitutes an effective strategy to overcome data scarcity and enhance generalization in blood vessel segmentation.