Development and validation of a deep learning model for automatic assessment of corneal endothelium

Aims To describe and validate a deep learning (DL) model for fully automatic assessment of corneal endothelium cell (CEC) parameters from specular microscopy images. Methods A total of 896 images were collected retrospectively from 440 subjects, including healthy and diseased corneas with varying image quality. Five DL models with different architectures were trained to extract eight key CEC characteristics, including cell density (CD), coefficient of variation (CV), pachymetry, average cell area (Ave), cell area standard deviation (SD), hexagonality coefficient (HEX), maximum (Max) and minimum (Min) cell area. Ground truth (GT) was derived from built-in software of the specular microscopy (Konan NSP-9900) with manual cell-centre annotation. An independent dataset comprising 100 images from 100 subjects was used for external validation. Results Among all tested models, the ResNet V.1 101 model achieved the best performance with the lowest mean squared error (MSE) of 0.24. The model demonstrated high accuracy in quantifying CEC parameters, particularly for CV, Ave, SD, Max and Min, with deviations ranging from 0.93% to 1.93%. The mean deviation was 3.10%±2.74% across all parameters. DL prediction and GT showed strong and significant correlations for CD, Ave, SD, Max and Min (r=0.89–0.99; all p<0.001), with excellent agreement for CD (intraclass correlation coefficient 0.964 (95% CI 0.944 to 0.976), p<0.001). External validation showed comparable predictive performance with a mean MSE of 0.26 and a mean deviation of 4.39%±2.60%. Conclusions The proposed DL model showed robust performance in quantifying eight key CEC parameters across specular microscopy images with varying corneal health conditions and image quality.