Abstract
Leakage in retinal angiography is a critical indicator for managing diseases like diabetic maculopathy and paediatric malarial retinopathy. This paper introduces a novel saliency-based framework for automatic leakage detection in fluorescein angiography (FA). The method utilizes a multi-scale superpixel approach combined with two specific saliency cues: intensity and compactness. By fusing these cues across different scales and applying graph-cut segmentation, the system identifies leaking regions with high precision. Experimental results on public datasets for malarial and diabetic retinopathy demonstrate that the proposed method performs at a level comparable to human experts and outperforms existing state-of-the-art automated techniques.
Introduction
In retinal diseases such as diabetic retinopathy (DR) and malarial retinopathy (MR), blood vessels often exhibit "leakage," where fluid escapes into the surrounding tissue. Clinicians visualize these leaks using Fundus Fluorescein Angiography (FA), where a fluorescent dye highlights active lesions as bright spots. Accurate identification of these leaks is vital for diagnosis and monitoring treatment efficacy. However, manual delineation is labor-intensive and prone to inter-grader variability. Automating this process is difficult due to the irregular shapes of leakages and the presence of confounding bright anatomical structures, such as the optic disc, or artifacts from uneven illumination.
Methodology
The authors propose a multi-stage pipeline that frames leakage detection as a saliency estimation problem, focusing on identifying regions that are most conspicuous relative to their surroundings.
- Multi-scale Superpixels: To account for varying leakage sizes, the image is divided into superpixels using the SLIC algorithm at three distinct scales (333, 666, and 1000 superpixels).
- Dual Saliency Cues:
- Intensity: A Kernel Density Estimator (KDE) is used to identify patches that are statistically brighter than the background.
- Compactness: This key innovation recognizes that true leakages are typically focal and compact, whereas noise or lighting artifacts are often diffuse. A spatial variance score is calculated for each patch to measure brightness concentration.
- Fusion and Refinement: Saliency maps are averaged across scales and then combined using pixel-wise multiplication. This ensures only regions that are both bright and compact are prioritized. The final map is thresholded and refined using Graph-Cut segmentation to ensure smooth boundaries, with anatomical masks applied to exclude the optic disc and blood vessels.
Experimental Results
The method was validated using two publicly available datasets: an MR dataset (25 images) and a DR dataset (24 images), with ground truth established by expert consensus.
The proposed method achieved state-of-the-art performance, matching human expert agreement levels. On the MR dataset, the algorithm reached a site-level sensitivity of 0.98 and a pixel-level accuracy of 0.91. For the DR dataset, it achieved an Area Under the Curve (AUC) of 0.86, significantly outperforming previous leading methods (e.g., Rabbani et al., which scored 0.80).
Discussion and Conclusion
The primary strength of this approach lies in the introduction of "compactness" as a discriminator, which effectively filters out pseudo-leakages caused by camera flash or anatomical features. The multi-scale fusion allows for the simultaneous detection of tiny punctate leaks and larger focal leaks.
While the results are promising, the study is limited by a relatively small sample size (~50 images). Future work will focus on validating the model on larger, more diverse datasets and automating parameter tuning to improve generalizability across different imaging conditions. This intensity and compactness framework holds potential for adaptation to other medical imaging tasks requiring the detection of focal lesions.
Summary updated