Evaluation of clinical consistency between deep learning algorithm-based ef-fective optical zone measurement after fully automatic corneal refractive sur-gery and traditional measurement methods
10.13389/j.cnki.rao.2025.0108
- VernacularTitle:基于深度学习算法的全自动角膜屈光手术后有效光学区测量与传统测量方法的临床一致性评价
- Author:
Yuhua ZHOU
1
;
Mengyang CHEN
;
Changtao YOU
;
Shuaifei LI
;
Lingling XU
;
Dongdong CHEN
;
Hongjie MA
;
Geng LI
;
Mingyang HU
Author Information
1. 450000 河南省郑州市,河南大学附属郑州爱尔眼科医院;475000 河南省开封市,河南大学淮河医院眼科
- Publication Type:Journal Article
- Keywords:
femtosecond laser-assisted in situ keratomileusis;
small incision lenticule extraction;
corneal topogra-phy;
artificial intelligence;
effective optical zone;
Bland-Altman analysis
- From:
Recent Advances in Ophthalmology
2025;45(8):629-634
- CountryChina
- Language:Chinese
-
Abstract:
Objective To investigate the diagnostic accuracy and clinical applicability of the Linknet-VGG16 deep learning algorithm for measuring the effective optical zone(EOZ)after corneal refractive surgery.Methods This single-center retrospective cohort study included 69 patients(69 eyes)who underwent femtosecond laser-assisted in situ kerato-mileusis(FS-LASIK)(34 eyes)or small incision lenticule extraction(SMILE)(35 eyes)at the Refractive Surgery Center of Affiliated Zhengzhou Aier Eye Hospital of Henan University from June 2023 to June 2024.Data from the right eyes of all patients were selected for statistical analysis.During the surgery,patients in the FS-LASIK group adopted the VisuMax fem-tosecond laser system combined with the Amaris 750S excimer laser system,while those in the SMILE group only used the VisuMax femtosecond laser system.A total of 276 Pentacam images were re-examined postoperatively.A Linknet segmenta-tion model based on the VGG16 encoder was constructed,and image normalization techniques were applied to accelerate model convergence.Model performance was assessed using accuracy,intersection over union(IoU),and the Dice coeffi-cient.The traditional EOZ measurement method based on corneal tangential curvature served as the reference standard.Bland-Altman analysis was conducted to evaluate consistency across all images and within each group,and the time effi-ciency of both methods was compared.Results Six representative medical image segmentation architectures(U-Net,U-Net++,DeepLabv3-ResNet50,DeepLabv3+-ResNet50,Unet-Densenet169,and Linknet-VGG16)were systematically evaluated.The Linknet-VGG16 model demonstrated superior performance over the other 5 models in pixel-level accuracy,IoU and Dice coefficient,which were 99.83％,99.48％and 99.74％,respectively.Although there was no significant differ-ence in accuracy and Dice coefficient between Linknet-VGG16 and U-Net models(whose accuracy was 99.82％and Dice coefficient was 99.72％),the inference speed of the U-Net model(62.46 ms)was 31.76％slower than that of the Linknet-VGG16 model(42.62 ms).The evaluation results of a clinically applicable comprehensive scoring model(weights:accura-cy 20％,IoU 20％,Dice coefficient 20％,speed 25％,model size 15％)showed that the Linknet-VGG16 model achieved a score of 88.01,surpassing other architectures(U-Net:86.29;DeepLabv3+-ResNet50:80.41;DeepLabv3-ResNet50:73.82;U-Net++:73.22;Unet-Densenet169:66.66).Bland-Altman analysis revealed that the mean difference of the 136 images in the FS-LASIK group was 0.01 mm[95％limits of agreement(LoA):-0.36 to 0.35 mm],with 96.3％of data points falling within the LoA.The mean difference of the 140 images in the SMILE group was-0.01 mm(95％LoA:-0.36 to 0.33 mum),with 95.7％of data points falling within the LoA.The mean difference of all 276 images was 0.00 mm(95％LoA:-0.36 to 0.34 mm),with 96.4％of data points falling within the LoA.These results indicated excellent consistency.The average measurement time per image using the traditional EOZ measurement method was 13.00 minutes,whereas the deep learning model required only 3.22 seconds.Conclusion The traditional EOZ measurement method based on corne-al tangential curvature exhibits good consistency with the fully automatic EOZ measurement method based on deep learning algorithms,achieving high image recognition accuracy.Additionally,the deep learning algorithm significantly reduces measurement time,compared with the traditional method based on corneal tangential curvature.
