ACCURATE CELL SEGMENTATION IN BLOOD SMEAR IMAGES BASED ON COLOR ANALYSIS AND CNN MODELS

Abstract

Nowadays, automated blood cell evaluation play a major role in the classification and diagnosis of diseases. Despite the many possible ways to segment blood cells, the recognition efficiency remains insufficient, especially when different cell types overlap. Also, one should not forget about the cells structure complexity. Image segmentation and image classification are the main stages of this problem. At the same time, segmentation of blood smear images is considered the most important stage in automated disease detection systems. Often cell segmentation in blood smear images is performed as a separate mapping for white blood cells and red blood cells. We propose another problem statement that uses the capabilities of supervised and unsupervised CNNs for the semantic segmentation of objects of different sizes and shapes. CNN has encoder-decoder architecture and builds a pseudo-color map. We tested several CNN models using different color spaces converting initial images from RGB to Lab, HSV and CMYK color spaces and obtained promising experimental results for several microscopic datasets such as CellaVision DM96, All-IDB and Blood Cell Detection.