Segmentation-based Retinal Image Fusion for Hypertension Prediction

Retinal image vessel segmentation or artery/vein segmentation tasks have been investigated for a long time, and hypertension prediction based on statistical data or body indicators has been learned well. However, the gap between retinal image segmentation and retinal image hypertension prediction still exists. In this paper, we bridge the gap by introducing a model-agnostic cross attention module in segmentation part and a semantic image fusion module in hypertension prediction part thus to form a novel segmentation-based pipeline for hypertension prediction. Specifically, the cross attention module adopts cross multiplication to attend encoder and decoder features, which enhances the artery/vein segmentation ability in uncertain region and border region in retinal image. Then we design a semantic image fusion module to fuse segmented artery/vein vessel image and original image as the classifier input to predict hypertension. The experimental results demonstrate that our model can efficiently predict hypertension, and we achieve 94.87% accuracy, 94.74% specificity, 95.00% sensitivity respectively on 393 retinal images from Kaggle ODiR5k dataset at : https://www.kaggle.com/ocular-disease-recognition-odir5k.