基于CNN和Transformer的睑板腺图像多粒度分割

打开文本图片集

关键词：睑板腺图像分割；多粒度分割;CNN;Transformer；医学图像处理中图分类号：TP391.41文献标识码：Adoi：10.37188/OPE.20253320.3299 CSTR：32169.14.OPE.20253320.3299

Abstract： To address the multi-stage processing and edge blurring issues in meibomian gland image segmentation，this paper designed an end-to-end multi-granularity segmentation algorithm.During the encoding phase，the TransUNet encoder architecture was adopted to eficiently extract shared features of the eyelid and glandular regions.In the decoding phase，a dual decoding path was employed to set up different decoder branches for the unique features of the eyelid and glandular regions.Meanwhile，for the glandular region，a multi-scale feature fusion module was designed，and a channel attention mechanism was incorporated into the skip connections. These optimizations improved edge accuracy，texture clarity，and shape contour，thereby effectively solving the problems of edge blurring and glandular adhesion.For the eyelid region，a standard decoder structure was used for segmentation prediction. Through experimental comparison with existing advanced segmentation methods，the proposed algorithm exhibits excellent performance in terms of the average acuracy for the upper and lower meibomian glands.Especially on the key indicators of mean Intersection over Union（IoU） and Dice Similarity Coeficient， it reaches 79.9% and 76.5% respectively，which are 3.2% and 5.3% higher than those of TransUNet. The algorithm in this paper can accurately segment the target regions of meibomian gland images，which can provide a basis for the auxiliary diagnosis of meibomian gland dysfunction.

Key words：meibomian gland image segmentation；multi-granularity segmentation；CNN；Transformer； medical image processing

1引言

医学图像分割作为计算机辅助检测流程的第一步，旨在识别感兴趣对象的轮廓或区域信息[]，其核心目标是将医学图像中的目标病变组织与无关组织分离2，精准勾画出感兴趣的区域，例如患者体内的不同组织结构、异常病变或解剖结构[3]。（剩余23016字）