科研之友
微信
新浪微博
Facebook
分享链接
摘要
Fully automated detection of longitudinal changes for retinal fundus images is a critical issue related to medical image analysis. Classical methods employ image differencing or image ratioing. However image differencing is susceptible to the illumination variations and image ratioing often fails in detecting the small lesions. This paper presents an unsupervised change detection method based on low-rank matrix decomposition. The proposed method models the anatomic structures with the illumination as the background and only longitudinal changes remain in the foreground. It avoids locating and segmenting anatomic structures, and detect the change features directly by decomposing the matrix composed of longitudinal image serial. The proposed method is also robust against the illumination variations through combining the intra-image correction with the inter-image normalization and linear interpolation. Extensive experimental results and comparisons demonstrate that the proposed method is effective and robust with illumination variations on detecting both the big lesions and the small lesions.
