Gland Segmentation In Pathological Images Based On Hybrid Feature Enhancement Network

The gland is an important structure in most organs and is responsible for secreting carbohydrates and proteins.Meanwhile,malignancy caused by the epithelium of the gland,also known as adenocarcinoma,is one of the most common cancers.Glandular morphology is especially important in determining the grade of cancer,which can be used to guide doctors in making treatment plans.Therefore,it is of great significance to accurately segment the shape of the gland in histopathological images,and it is also the premise of carrying out the pathological image-assisted diagnosis.In general,artificial segmentation is more accurate than automatic segmentation.However,due to the large size of digital pathological images,manual segmentation is extremely time-consuming and expensive.Therefore,it is valuable and meaningful to study the method of automatic segmentation,which can reduce the work intensity of pathologists.In the context of rapid gland segmentation by computer-assisted technology,accurate gland segmentation remains a challenge,such as adhesion between adjacent glands and large morphological differences between benign and malignant glands.in-depth research on automatic gland segmentation,which mainly includes:(1)This paper proposes a Hybrid Feature Enhancement Network(HFE-NET)to enhance global information,local information,multi-scale information,and boundary information.The network includes several Multi-Scale Local Feature Extraction blocks(MSLFEB)and Global Feature Enhancement blocks(GFEB).MSLFEB can extract multi-scale features through different sizes of receptive fields,reduce the loss of local information,and effectively alleviate the adhesion problem of the gland.The GFEB module enhances the global semantic information and transmits the underlying features to the decoder.By integrating the information of high and low dimensions,the gland segmentation of the target region is more accurate.(2)Considering the class imbalance in the gland data set and the variable shape of the gland,this paper designed a class balance Variance Loss Function(Focal and Variance Loss Function(FV-Loss).The FV-loss effectively alleviates the problem of quasi-imbalance between the gland and the background and restricts the pixels in the instance of the same gland to avoid voids in the instance of the gland.(3)We conducted experiments on two open datasets with significantly different morphologies of the glands and compared them with existing automatic segmentation methods.The results show that HFE-NET can effectively enhance the global,local,and boundary information,and alleviate the adhesion problem of adjacent glands.Meanwhile,FV-loss reduces the false positives of the model.Compared with the existing methods of gland segmentation,HFE-NET is better than most of the latest methods in many indicators,and the model is more generalized.