Manipulating Retinal OCT data for Image Segmentation based on Encoder-Decoder Network

Medical image analysis is consistently being researched in the computer vision in that it captures potential symptoms and enables more delicate diagnosis of patients. Based on the development of medical equipment such as optical coherence tomography(OCT) and magnetic resonance imaging(MRI), it is possible to analyze medical data with clearer and higher resolution than before. However, there are still many data that have limitations in manually diagnosis by human. Moreover, identifying the extent of the damaged retinal layer also remains one of the most challenging tasks since the damaged layer not only contains too many invisible layers, but it is too small. Normal OCT data has smooth layers while age-related macular degeneration(AMD) or diabetic macular edema(DME), which are classified as abnormal, has layers that are damaged by bleeding. The precise regional classification is required for the diagnosis and prescription of the damaged layers and a new approach to effectively training an irregular layer of abnormal data is also needed. Hence, this paper proposes an OCT data manipulation method as a preprocessing step to improve training boundaries of regional layers. The preprocessed data were generated by manual range using the proposed method and applied to the encoder-decoder networks, SegNet and Unet. The experiment shows that the preprocessed datasets were trained much faster than the original and the optimized range was also confirmed through comparison the results of preprocessed dataset by each range.