We propose a deep-learning based annotation-efficient framework for vessel detection in ultra-widefield (UWF) fundus photography (FP) that does not require de novo labeled UWF FP vessel maps. Our approach utilizes concurrently captured UWF fluorescein angiography (FA) images, for which effective deep learning approaches have recently become available, and iterates between a multi-modal registration step and a weakly-supervised learning step. In the registration step, the UWF FA vessel maps detected with a pre-trained deep neural network (DNN) are registered with the UWF FP via parametric chamfer alignment. The warped vessel maps can be used as the tentative training data but inevitably contain incorrect (noisy) labels due to the differences between FA and FP modalities and the errors in the registration. In the learning step, a robust learning method is proposed to train DNNs with noisy labels. The detected FP vessel maps are used for the registration in the following iteration. The registration and the vessel detection benefit from each other and are progressively improved. Once trained, the UWF FP vessel detection DNN from the proposed approach allows FP vessel detection without requiring concurrently captured UWF FA images. We validate the proposed framework on a new UWF FP dataset, PRIME-FP20, and on existing narrow-field FP datasets. Experimental evaluation, using both pixel-wise metrics and the CAL metrics designed to provide better agreement with human assessment, shows that the proposed approach provides accurate vessel detection, without requiring manually labeled UWF FP training data.

While recent advances in deep learning have significantly advanced the state of the art for vessel detection in color fundus (CF) images, the success for detecting vessels in fluorescein angiography (FA) has been stymied due to the lack of labeled ground truth datasets. We propose a novel pipeline to detect retinal vessels in FA images using deep neural networks (DNNs) that reduces the effort required for generating labeled ground truth data by combining two key components: cross-modality transfer and human-in-the-loop learning. The cross-modality transfer exploits concurrently captured CF and fundus FA images. Binary vessels maps are first detected from CF images with a pre-trained neural network and then are geometrically registered with and transferred to FA images via robust parametric chamfer alignment to a preliminary FA vessel detection obtained with an unsupervised technique. Using the transferred vessels as initial ground truth labels for deep learning, the human-in-the-loop approach progressively improves the quality of the ground truth labeling by iterating between deep-learning and labeling. The approach significantly reduces manual labeling effort while increasing engagement. We highlight several important considerations for the proposed methodology and validate the performance on three datasets. Experimental results demonstrate that the proposed pipeline significantly reduces the annotation effort and the resulting deep learning methods outperform prior existing FA vessel detection methods by a significant margin. A new public dataset, RECOVERY-FA19, is introduced that includes high-resolution ultra-widefield images and accurately labeled ground truth binary vessel maps.

Wide-field fluorescein angiography (FA) images are commonly used in ophthalmology to assess longitudinal changes in retinal vasculature, specifically, non-perfusion. Current practice relies on manual qualitative comparisons between images taken at successive clinic visits, a few months apart. Objective quantitative assessments, although desirable for evaluating disease progression and treatment, are impractical to perform manually and challenging for image analysis because of the changes in the capture viewpoints and temporal imaging variations seen as the FA dye injection perfuses the retina. We propose a methodology for quantifying retinal non-perfusion by automated analysis of the FA images captured during successive clinical visits. Blood vessels are first detected in the image from each visit. The vascular networks in FA images are then precisely registered to obtain a co-aligned pair via parametric chamfer matching under polynomial transformation, a process that explicitly allows for increase or decrease in perfusion. Changes in perfusion are then quantified by identifying the common and distinct regions in co-aligned image pairs. The proposed framework is tested on FA images that are manually annotated by an ophthalmologist to provide ground truth binary vessel masks and to identify vasculature changes. Results indicate that the proposed method provides assessments of vasculature changes that are in good agreement with the ophthalmologist-provided annotations.

Retinal blood vessel detection is a crucial step in automatic retinal image analysis. Recently, deep neural networks have significantly advanced the state of the art for retinal blood vessel detection in color fundus (CF) images. Thus far, similar gains have not been seen in fluorescein angiography (FA) because the FA modality is entirely different from CF and annotated training data has not been available for FA imagery. We address retinal vessel detection in wide-field FA images with generative adversarial networks (GAN) via a novel approach for generating training data. Using a publicly available dataset that contains concurrently acquired pairs of CF and fundus FA images, vessel maps are detected in CF images via a pre-trained neural network and registered with fundus FA images via parametric chamfer matching to a preliminary FA vessel detection map. The co-aligned pairs of vessel maps (detected from CF images) and fundus FA images are used as ground truth labeled data for de novo training of a deep neural network for FA vessel detection. Specifically, we utilize adversarial learning to train a GAN where the generator learns to map FA images to binary vessel maps and the discriminator attempts to distinguish generated vs. ground-truth vessel maps. We highlight several important considerations for the proposed data generation methodology. The proposed method is validated on VAMpIRE dataset that contains high-resolution wide-field FA images and manual annotation of vessel segments. Experimental results demonstrate that the proposed method achieves an estimated ROC AUC of 0.9758.