Global Contrastive Learning Medical Images Segmentation Method For Class-imbalance Problem

Class imbalance issue is commonly observed in real image datasets,which lowers the accuracy of visual classification or segmentation tasks.In the domain of clinical healthcare,the segmentation accuracy of medical images is directly related to clinical diagnosis and treatment.In recent years,computer vision technology has achieved great progress in clinical applications.However,in existing classification segmentation methods,it usually focuses only on the information within a single image,such that the accuracy is impaired in class unbalanced scenarios.To address class-imbalance and poor segmentation details in retinal OCTA images,the contributions of this thesis are as follows:1.This thesis summarizes the contrastive learning structure,and proposes hard anchor sampling and semi-hard sampling strategies.Experiments demonstrate the efficiency of the proposed method in this paper.The proposed method is generalizable in the image segmentation task with class imbalance and can be migrated to other image segmentation tasks.2.The two-branch network architecture COSNet is proposed for retinal image segmentation in different depths.The method includes a feature extraction module,a contrastive learning module and a fine-tuning module to construct a global contrast learning loss,and finally,it is able to implement adaptive clustering fine-tuning for location information.The experiments prove that COSNet shows some advantages in different metrics of segmentation.3.Based on the above two methods,an online platform for artificial intelligence medical image segmentation is developed with functions such as registration and login,uploading dataset,segmentation test,image segmentation and history recording.