A Copyright- and Privacy-Protected Diabetic Retinopathy Diagnosis Network

This paper proposes a copyright- and privacyprotected diabetic retinopathy (DR) diagnosis network. In the network, DR lesions are automatically detected from a fundus image by firstly estimating non-uniform illumination of the image, and then the lesions are detected from the balanced image by using level-set evolution without re-initialization. The lesions are subsequently marked by using contours. The lesion-marked fundus image is subsequently shared for intra or inter hospital network diagnosis with copyright and privacy protection. Watermarking technique is used for image copyright protection, and visual encryption is used for privacy protection. Sign scrambling of two dimensional (2D) discrete cosine transform (DCT) and one dimensional (1D) DCT is proposed for lesion-marked fundus image encryption. The proposed encryption methods are compared with other transform-based encryption methods, i.e., discrete Fourier transform (DFT) amplitude-only images (AOIs), DCT AOIs, and JPEG 2000-based discrete wavelet transform (DWT) sign scrambling which were proposed for image trading system. Since the encryption is done after DR diagnosis, contours used for DR marking must also be visually encrypted. The proposed encryption methods are effective for strong-edge images that are suitable for lesion-marked fundus images, while random sign-based JPEG 2000, DFT AOIs, and DCT AOIs encrypt the images imperfectly. Moreover, the proposed methods are better in terms of image quality. In addition, watermarking performance and compression performance are confirmed by experiments.