Retinal microaneurysms detection using adversarial pre-training with unlabeled multimodal images

The detection of retinal microaneurysms is crucial for the early important diseases such as diabetic retinopathy. However, these lesions in retinography, most widely available imaging modality, remains a very challenging task. This mainly due to tiny size and low contrast images. Consequently, automated usually relies on extensive ad-hoc processing. In this regard, although can be more easily detected using fluorescein angiography, alternative modality invasive not adequate regular preventive screening. work, we propose novel deep learning methodology that takes advantage unlabeled multimodal image pairs improving retinography. particular, adversarial pre-training consisting prediction angiography from retinography generative networks. allows about retina without any manually annotated data. Additionally, also approach heatmap regression, which an efficient precise localization multiple microaneurysms. To validate analyze proposed methodology, perform exhaustive experimentation different public datasets. provide relevant comparisons against state-of-the-art approaches. results show satisfactory performance proposal, achieving Average Precision 64.90%, 31.36%, 33.55% E-Ophtha, ROC, DDR Overall, outperforms existing alternatives while providing straightforward method effectively applied raw unprocessed