肺组织术中冷冻切片AI虚拟染色的诊断价值OA

Objective To overcome the cumbersome and time-consuming procedures of chemical staining for in-traoperative frozen sections,as well as inconsistencies caused by variable experimental conditions,we constructed a deep learning-based artificial intelligence(AI)virtual staining model for intraoperative frozen sections of lung tissue.This model aims to achieve accurate transformation from label-free frozen sections to standard hematoxylin-eosin(HE)stained images,thereby improving the timeliness of intraoperative frozen pathological diagnosis.Methods A mix-domain sinkhorn optimal-transport bridge(MSOTB)model was established,which formulated the unpaired stain trans-fer task as a stepwise distribution transport process driven by a stochastic bridge.The model comprised four core mod-ules:a time-conditioned generator(G),discriminator(D),energy network(E),and feature extract network(F).A joint optimization strategy was adopted,integrating adversarial learning,bridge process consistency constraint,Sinkhorn-based OT-PatchNCE loss,and mix-domain contrastive learning to balance stylistic alignment and structural fidelity.A total of 50 consecutive lung tissue specimens were collected,and 30 000 images for model training and 1 000 paired images for validation were generated therefrom.Results On the test set,the MSOTB model outperformed baseline methods(including contrastive unpaired translation,mix-domain contrastive learning,and unpaired neural Schrödinger bridge)in three key metrics:Fréchet inception distance(FID=34.7),learned perceptual image patch similarity(LPIPS=0.21),and structural similarity index measure(SSIM=0.72).Pathological diagnostic evaluation showed that the diagnostic concordance between AI virtual staining and manual HE staining reached 94.80%,with a sensitivity of 94.74%.Ablation studies verified that each module contributed significantly to the model performance.Conclusion AI-based virtual HE staining can assist intraoperative frozen pathological diagnosis for most pulmonary diseases and exhibits promising application potential.However,it has certain limitations in diagnosing rare lesions such as mucinous tumors and cartilaginous tumors.Further optimization and validation based on multi-center studies with larger cohorts are warranted.