病理图像分割及评价方法研究

In medical diagnosis, information such as the morphological characteristics of the nucleus can be obtained by segmenting pathological images and can be used to grade cancer and evaluate the effect of treatment. The difficulty of nucleus segmentation is related to inhomogeneous cell staining and the morphological characteristics of dense bulk nuclei. In this study, the threshold segmentation and region segmentation models provided by ImageJ image processing software were used for kernel segmentation, and compared with the U-Net model segmentation, the results showed that the evaluation indexes of Acc, F1, MioU and other indicators of the U-Net nucleus segmentation algorithm reached 91.2%, 80.9%, and 78.5%, respectively, and the nucleus segmentation performance was stable, and the robust robustness and adaptability of the U-shaped convolutional network made automatic nuclei segmentation possible.