Proceedings Article | 22 October 2021
KEYWORDS: Image segmentation, Data modeling, Computer programming, Feature extraction, RGB color model, Performance modeling, Image enhancement, Kidney, Tissues, Convolution
The classification and detection of glomeruli of kidney tissue is a key process for correct diagnosis of diseases in renal pathology. However, it is still a big problem to perform comprehensive and accurate glomerular ultramicroscopic pathological diagnosis based on high-resolution whole-field digital slices (whole Slide Image). The reason lies in the grayscale image texture corresponding to the glomerular ultrastructure. Complicated, there are many types of related lesions, and it is difficult to identify and judge subtle pathological changes. The traditional semantic segmentation model cannot achieve the ideal segmentation effect. Based on this, this paper proposes a semantic segmentation method FEU-Net (FCN-Efficient U-Net) for pathological slices, which improves the accuracy of glomerular region segmentation and realizes end-to-end segmentation. FEU-Net uses EfficientNet, which has undergone transfer learning, as the encoder part to enhance image feature extraction capabilities. The decoder uses U-Net to promote the fusion of deep and shallow features while reducing the amount of network parameters, and redesign the convolution module to improve the gradient transfer capability. Compared with other classic methods in the SEED data set, it verifies that some classic models are difficult to fit in this segmentation task. At the same time, experiments show that modifying the feature extraction method can greatly improve the results. Among them, the modified method in this article is in the accuracy of segmentation. This is an increase of 18.968.96% over the original U-Net. At the same time, this article conducted ablation experiments on the ZENODO data set and the mendeley data set, verifying that each module in the improved algorithm helps to improve the segmentation effect of pathological slices. In the SEED data set, the FEU-Net method in this paper improves the Dice coefficient, accuracy rate and recall rate by 5.175.17%, 2.72.7%, 3.693.69%, 4.084.08% respectively compared with the benchmark model; in the BOT data set, The three indicators of the method in this paper have increased by 0.47%, 0.060.06%, 4.304.30%, and 6.086.08% respectively. The FEU-Net proposed in this paper improves the accuracy of segmentation of the pathological slices of gastric cancer and has good generalization performance.
