Multi-input 2-dimensional deep belief network: diabetic retinopathy grading as case study

The most important action in treating diabetic retinopathy is early diagnosis and its progression degree. This paper presents a two-dimensional Deep Belief Network based on Mixed-restricted Boltzmann Machine capable of receiving multiple two-dimensional inputs. Using multiple inputs provides more appropriate prior information for learning. In this proposed method, the image is transferred to the HSV color space and then the 3D color image is converted to a 2D matrix using a weighted mean. This weighted mean is calculated based on the entropy criterion. The resulting two-dimensional matrix is not in pixel and is merely a raw description of the image. The local, regional and global descriptions are extracted from this matrix and provided for the network. The proposed deep network automatically extracts the appropriate features to determine the progression degree of diabetic retinopathy by the network. Window by window image processing can overcome one of the basic problems of image classification, i.e. the small number of labeled data. Experiments showed that the proposed method is superior when compared to other methods.