Momaku: A Retinal Image Annotation Platform

Efficient and accurate data annotation is essential for developing machine learning models in medical imaging, particularly for retinal image analysis. However, existing annotation tools often lack the specificity required for detailed retinal structures, especially when handling challenging images with less-than-optimal visual quality like premature retinal images. To address this, we introduce Momaku, a web-based, freely-available platform primarily designed for annotating premature retinal fundus images but potentially applicable to other contexts. Momaku offers a clinician-friendly interface, precise drawing tools, and integration with vascular feature extraction libraries to enhance annotation value and accuracy. In a preliminary study, we utilized Momaku to remove ground-truth annotation artifacts from an existing public database for retinal vessel segmentation purposes. We then trained two standard vessel segmentation models using raw and corrected annotations and evaluated performance using overlap and topological correctness metrics. Experimental results demonstrate that improved annotation quality can lead to better segmentation performance, validating the need for specialized annotation platforms that can enable efficient quality control and ground-truth correction on medical segmentation tasks.