Abstract: Aiming at the problem that it was difficult to segment adhesive leukocyte accurately, we proposed an adhesive leukocytes segmentation algorithm based on deep learning. Firstly, the color space of the blood cell microscopic images of patients with acute lymphoblastic leukemia was transformed
 from RGB to HSV, in order to filter out red blood cells and extract leukocytes. Secondly, for the adhesive leukocytes in extraction results, the cell border was set as the third class, in addition to foreground and background. During the training process of deep learning segmentation model, a weighted cross-entropy loss function based on class weight was introduced to make the model learn more cell border features. The experimental results show that using the proposed method to segment the leukocytes in the dataset ALL_IDB1 can achieve an accuracy of 95.19%.

Key words: adhesive leukocyte segmentation, color space transformation, weighted loss function