Background Generalized knee tissue segmentation,such as cartilage and meniscus in magnetic resonance imaging(MRI),plays a vital role in the clinical assessment of knee osteoarthritis(OA).However,domain variability bet...Background Generalized knee tissue segmentation,such as cartilage and meniscus in magnetic resonance imaging(MRI),plays a vital role in the clinical assessment of knee osteoarthritis(OA).However,domain variability between MRI datasets poses a significant challenge for the application of robust segmentation methods in real-world clinical settings.Existing unsupervised domain adaptation(UDA)approaches,which rely on one-to-one assumptions between the source and target domains,often fail to preserve knee tissues such as cartilage and meniscus,which are critical for OA diagnosis in diverse clinical settings.Methods We propose a source-independent segmentation approach tailored for multi-domain knee MRI datasets.Our method emphasizes knee tissue regions to reduce domain gaps and label inconsistencies.By introducing a stepwise adaptation strategy,segmentation performance was refined progressively from intermediate domains to the final target domain.Pseudo-label attention mechanisms were integrated into the adaptation pipeline,enabling iterative fine-tuning of domain-specific segmentations while leveraging unidirectional generative adversarial networks to enhance tissue-specific adaptation.This iterative training process ensures the generation of reliable pseudo-labels,thereby improving segmentation accuracy in diverse clinical MRI datasets.Results We demonstrated the effectiveness of our approach on the OA initiative dataset as the source domain and self-collected,T1-weighted fast field echo(T1FFE)as the intermediate domain and three-dimensional fast spin echo(3D FSE)as the final target domain.Our method achieved an average dice scores of 0.8701 and 0.7990 for source and target domains,respectively,surpassing the typical UDA methods explored in our experiments.Conclusion The experiments conducted on clinical MRI data,spanning OA severity from healthy knees to KL Grades 1–4,validated the effectiveness of the proposed domain adaptation method in precise segmentation of the cartilage and meniscus.展开更多
基金supported by a grant from the Innovation and Technology Commission of the Hong Kong SAR(Project MRP/001/18X)a grant from the Research Grants Council of the Hong Kong SAR(UGC/FDS24/E18/22).
文摘Background Generalized knee tissue segmentation,such as cartilage and meniscus in magnetic resonance imaging(MRI),plays a vital role in the clinical assessment of knee osteoarthritis(OA).However,domain variability between MRI datasets poses a significant challenge for the application of robust segmentation methods in real-world clinical settings.Existing unsupervised domain adaptation(UDA)approaches,which rely on one-to-one assumptions between the source and target domains,often fail to preserve knee tissues such as cartilage and meniscus,which are critical for OA diagnosis in diverse clinical settings.Methods We propose a source-independent segmentation approach tailored for multi-domain knee MRI datasets.Our method emphasizes knee tissue regions to reduce domain gaps and label inconsistencies.By introducing a stepwise adaptation strategy,segmentation performance was refined progressively from intermediate domains to the final target domain.Pseudo-label attention mechanisms were integrated into the adaptation pipeline,enabling iterative fine-tuning of domain-specific segmentations while leveraging unidirectional generative adversarial networks to enhance tissue-specific adaptation.This iterative training process ensures the generation of reliable pseudo-labels,thereby improving segmentation accuracy in diverse clinical MRI datasets.Results We demonstrated the effectiveness of our approach on the OA initiative dataset as the source domain and self-collected,T1-weighted fast field echo(T1FFE)as the intermediate domain and three-dimensional fast spin echo(3D FSE)as the final target domain.Our method achieved an average dice scores of 0.8701 and 0.7990 for source and target domains,respectively,surpassing the typical UDA methods explored in our experiments.Conclusion The experiments conducted on clinical MRI data,spanning OA severity from healthy knees to KL Grades 1–4,validated the effectiveness of the proposed domain adaptation method in precise segmentation of the cartilage and meniscus.
