基于CNN、Transformer和Mamba模型的慢性根尖周炎病变自动分割性能比较OA

Objective To compare the performance differences of three major deep learning architectures,namely convolu-tional neural network(CNN),Transformer and Mamba,in the automatic segmentation task of chronic apical periodontitis(CAP)lesions in periapical radiographs.Methods A CAP-APX500 dataset was constructed by retrospective single-center collection of periapical radiographic data from 500 CAP patients at the Department of Stomatology of the First Medical Center of Chinese PLA General Hospital.Under a unified data partitioning scheme,training strategy and experimental environment,CNN models of U-Net,ResUNet++and ColonSegNet,Transformer models of Swin-UNet,PVT-CASCADE and PVTFormer and Mamba models of Mamba-UNet,VM-UNet and VM-UNetV2 were compared in terms of the performa-nce for automatic CAP segmentation of periapical radiographs.A comprehensive evaluation was conducted using metrics such as the Dice coefficient,intersection over union(IoU)and Hausdorff distance(HD).Results Transformer models behaved the best in all the models,followed by Mamba models and traditional CNN models in order.The PVTFormer model achieved the highest segmentation accuracy on the test set,with an average Dice coefficient of 0.657,an average IoU of 0.518,a precision of 0.757,a recall of 0.648,an F2 score of 0.647 and an HD of 3.39,all of which outperformed the other models(P＜0.01).Conclusion Deep learning models such as CNN,Transformer and Mamba segment CAP lesions on periapical radiographs effectively,of which,the PVTFormer model performs best in terms of segmentation accuracy and robustness,enabling effective segmentation of CAP lesion areas on periapical radiographs and providing a reliable solution for the computer-aided diagnosis and lesion assessment of CAP.[Chinese Medical Equipment Journal,2026,47(6):11-19]