TransUNet-based segmentation of knee meniscus in medical imaging
CAI Yi-bin
WANG Yu-ling
Abstract：Objective To improve the segmentation accuracy and boundary alignment of the meniscus in knee MRI images,particularly in four key anatomical regions(anterior/posterior horns of lateral/medial menisci).This study proposes an improved TransUNet-based high-precision automatic meniscus segmentation model.Methods Using the publicly available MRNet dataset from Stanford University,we constructed a finely annotated meniscus segmentation subset labeled by experienced radiologists as training and evaluation samples.The dataset was split into an 8:2 ratio for training and testing.Through transfer learning,we developed an ECMA-TransUNet model,incorporating an Efficient Channel and Multi-Scale Attention(ECMA)module embedded into the CNN encoder stages and three skip connection paths of the TransUNet decoder.This model was applied to segment the four meniscal horns.Results The proposed model achieved the following segmentation performance across the four anatomical regions:lateral anterior horn(DSC=87.43%,HD=0.5678),lateral posterior horn(DSC=93.15%,HD=0.8455),medial anterior horn(DSC=91.48%,HD=0.7551)and medial posterior horn(DSC=94.00%,HD=0.7407).All metrics met clinical usability standards.Conclusions The proposed ECMA-TransUNet demonstrates significant advantages:(1)The introduced attention module enhances boundary alignment and small-structure recognition;(2)The anatomical segmentation accuracy meets clinical diagnostic requirements,providing reliable support for intelligent diagnostic assistance systems.
Keywords：Knee jointMeniscusTransUNetImage segmentation
Publication Date：2026-02-19
Online Publishing Date：2026-03-09(First online date of this platform, not the publication date of the document)
Pages：6( 97-102 )