Abstract: The number of image samples of rare fundus diseases is small, which is difficult to meet the needs of deep network training. A recognition model with MFPD(Multi-representation Fused with Prior Distributions) for rare fundus diseases is proposed. Based on the pre-training model, fine tune the embedded model to obtain the feature embedded prior distribution, map the embedded features to different spaces, and extract the image features from different perspectives. On the basis of cross entropy loss, divergence loss is added to increase the difference of different perspective features, make efficient use of rare disease image information, and reduce the impact of small sample size on the model. The experiment uses the OPHDIAT(Ophtalmologie-Diabˋete-Te′le′me′decine) fundus image dataset to compare this method with other methods. The experimental results show the effectiveness of this method.

Key words: prior distribution')">

prior distribution, rare disease, embedded features, multiple representations, divergence loss