The Instance Segmentation Of Leukocyte Image Based On Deep Learning

Current blood smear examination mainly relies on pathological analysis by hematologists,which is time-consuming and subjective.With the continuous breakthrough of deep learning in the field of image,it provides a new idea for the analysis of leukocyte images.But most methods can only achieve a single task in segmentation or classification,while there are few algorithms achieving the segmentation and classification of multiple leukocytes in the whole cell image simultaneously,and it still can not handle effectively difficulties situations such as boundary blur,uneven staining and nuclear intra-class variability.To address these challenges,our contributions are as follows:Firstly,for the questions about boundary blur and nuclear intra-class variability,we propose an end-to-end framework based on U-Net for instance segmentation,integrating convolutional block attention modules(CBAM)and guided attention modules,which reflects the attention-inspired mechanism for integrating the local features with global dependencies in the small shapes of leukocyte objects.It can simultaneously segment the leukocytes accurately and classify the leukocytes into three sub-types by perceiving the shape and boundary issues of multiple leukocyte objects.In addition,we validate our method through extensive experiments on our own established wright stained leukocyte dataset.For the difficulties about the results of stained leukocyte images' instance segmentation heavily rely on the quality of staining and the datasets need to be labeled manually,an end-to-end framework is put forward which can segment the unstained leukocytes accurate by unsupervised learning.In addition,to further enhance the model,we combine image pyramid with the input of segmentation network to fusion image feature information at multiple scales.To achieve a higher accuracy so that leukocyte can be classified,adversarial learning is adopted in the output space to consider spatial similarities between the source and target domains.Finally,it regards wright stained leukocyte data as the source domain dataset and unstained leukocyte data as the target domain dataset,comparative experiments and ablation experiments are carried out to verify the effectiveness of the model.Extensive experiments show that the proposed methods perform favorably against the state-of-the-art methods in segmentation and classification.