Cell Segmentation Algorithm of Fully Connected U-Net ++ Network

With the continuous emergence and wide application of new medical imaging technologies, doctors can easily use them to provide key and reliable information for the diagnosis of lesions in real time. At the same time, with the continuous development of computer vision, deep learning has derived many model structures in the field of medical image segmentation. Among them, U-net network structure has become one of the important segmentation models in the field of medical image because of its clever skip connection, which can retain complete features while locating clear boundaries. Algorithms based on deep learning often rely on the training of massive data sets. In the field of cell images, dataset calibration is very difficult. Especially in multi-cell images, the position and boundary of nucleus and cytoplasm are ambiguous, and can not be correctly calibrated even by a large amount of manpower. And from the perspective of imaging principles, medical image data is complex and not informative enough. Based on the comparison between the improved fully connected U-net++ network and FCN, SegNet, U-Net and other network models, this paper proposes a data augmentation method for the selection and splicing of cell sample data from the perspective of data sampling and augmentation, which can effectively transform the weak supervised data containing part of the label into strong supervised data for training.