Advancing Retinal Vessel Recognition via Deep Learning with Limited Annotation
Hu, Dewei
0000-0001-7203-4136
:
2024-06-17
Abstract
As the revolutionary advancement in computational power has facilitated the direct manipulation of vast amounts of data, researchers have explored many algorithms capable of discerning the intrinsic patterns within this data. Such learning-based techniques have profoundly influenced the field of medical image analysis, despite the lack of human annotated ground truth hindering their broader application. In my research, I propose three solutions to address this challenge. First, I deploy a diffusion probabilistic model to reduce speckle noise in retinal optical coherence tomography (OCT) in an unsupervised manner. This approach significantly improves image quality, making retinal layers and vessels easier to capture. Secondly, I present a novel synthetic model that generates enhanced angiograms, allowing the binary vessel map to be acquired with simple thresholding. This idea extends to an unsupervised 3D vessel segmentation algorithm for retinal OCT angiography (OCT-A). By leveraging the inherent uncertainty of stochastic gradient descent, the model synthesizes diverse pseudo-modalities with consistent vessel morphology. For the third solution, I investigate domain generalization (DG) methods that enable the segmentation model to work on unseen datasets. I observe that the tubular shape of vessels remains invariant despite changes in image distribution. I propose an explicit shape representation of retinal vessels by forming a tensor field with Hessian eigenvectors, analogous to the diffusion tensor imaging used for delineating neural fibers. A transformer model is then trained on this handcrafted tensor field to segment vessels. To enhance model robustness and computational efficiency, an implicit shape representation is learned from the synthetic pseudo-modalities using contrastive learning. In addition, I also explore the domain adaptation (DA) methods and present to train the segmentation network with synthetic images generated by semantic conditional diffusion probabilistic model. In conclusion, the proposed algorithms facilitate deep models in recognizing retinal vessels with limited manual labels, advancing the field of medical image analysis despite the challenges posed by sparse annotations.