research design
This study was a retrospective diagnostic cross-sectional study from the medical records and imaging archives of Osaka University Hospital (OUH). This study followed the principles outlined in the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of Osaka University Hospital (identifier: 19492-T1). For children whose figures included photographs, informed consent was obtained from the parents or legal guardians of all child participants both for participation in the study and for publication of information and images in open access publications. rice field.
data set
In this cross-sectional study, we collected 65,534 periocular photographs of 8,125 patients from the OUH Ophthalmological Image Archive from November 2000 to September 2021. Images were carefully reviewed by multiple ophthalmologists and images with characteristic facial features were collected. Childhood glaucoma (corneal and/or ocular opacification and/or enlargement) was included in the glaucoma group. As a control group, primary gaze photographs of patients aged 0-10 years were selected among patients without childhood glaucoma. Because many pediatric glaucoma patients had strabismus, the control group also included patients with esotropia and exotropia in the same proportion as the glaucoma group. Age and gender distributions were adjusted.
Exclusion criteria were eyelid disease, hereditary facial malformations, and corneal disease. We also excluded photos with a tilted face, photos with glasses, and photos after eyelid or strabismus surgery. All photographs were taken at his 1 m distance from the subject using a commercial camera (D850; Nikon Inc., Tokyo, Japan) equipped with a ring flash. The original image was in Jpeg format and was 2034 x 1536 pixels in size.
Experimental device
Our methodology used a cross-validation approach with five stratified groups and one subject exclusion. Briefly, all images of a single subject with glaucoma or control group were first used as the test set. The remaining photographs were then randomly split into training and validation sets with a ratio of 80:20, and the proportions of esotropia, orthotropia, and exotropia were the same (1:5:4 ). Images of a single subject did not span multiple datasets (Figure 3).
Illustration of image dataset acquisition and preprocessing for childhood glaucoma detection using deep learning.
I implemented a deep learning algorithm using the PyTorch framework (version 1.10.0). The model was trained on Windows 10 Pro (version 21H2, build 19044.2486) operating system (Microsoft Corporation, Redmond, WA, USA) with Intel Xeon E-2276M CPU, 32 GB RAM, and NVIDIA Quadro RTX 5000 16 GB GPU. rice field. This research was conducted using the Windows 10 operating system. Algorithm performance was measured using diagnostic parameters such as accuracy, sensitivity, specificity, F value, and area under the receiver operating curve. We then compared the diagnostic accuracy of the deep learning algorithm with two glaucoma ophthalmologists and two pediatric ophthalmologists.
deep learning algorithm
We applied transfer learning based on a DCNN architecture pretrained on ImageNet (1,000 object categories with over 1 million images).DCNN applied was RepVGG-A231. To take advantage of the pre-trained DCNN model, the images were formatted as squares with one edge on the long side and black filled squares with a size of 250 × 250 pixels (Figure 4). Random horizontal flip, random rotation (± 1 degree), random resize (× 0.8 to × 1.1), and 224 × 224 × 3 for tuning the pretrained network to increase the number of images Data augmentation, such as trimming to , has been performed. Images were then normalized and standardized based on ImageNet values to improve uniformity. The training process utilized a randomization technique to allow the model to learn more image variations. For training, the Adabelief optimizer was selected for training based on its fast convergence and optimal performance as determined by pilot evaluations including SGD, RMSprop, Adam, AdamW, and Adabelief.32. After evaluating the impact of different learning rates (0.1, 0.01, 0.001, and 0.0001), the learning rate was set to 0.001. The network was trained for 200 epochs using a batch size of 16 and early stopping was applied if the validation loss did not decrease for 10 epochs (Table 5).
Structure of a convolutional neural network (RepVGG-A2). The input size of the network is 224×224.
Performance interpretation and statistics
All statistical analyzes were performed using the scikit-learn library for Python (version 3.8.6, Python Software Foundation, Beaverton, OR). Welch’s His t-test was used to compare the ages of glaucoma patients and controls. A chi-square test was used to compare the distribution of gender and eye position between the two groups. Statistical significance was defined as a p-value less than 0.05.
We evaluated the performance of the DCNN algorithm by determining the area under the receiver operating characteristic (ROC) curve (AUC). Diagnostic performance was evaluated for each image by calculating accuracy, sensitivity, specificity, positive predictive value, F value, and AUC. The F value was defined as (2 × sensitivity × positive predictive value)/(sensitivity + positive predictive value).33. In addition, we evaluated prediction performance when we ensembled 5-fold prediction results and defined 1-fold or more positive results as positive. For each analysis, the classification threshold was fixed at 0.5.
Gradient-weighted class activation mapping (Grad-CAM) was used to visually illustrate the deep learning algorithm.34.
Comparing Deep Learning Algorithms and Human Inspectors
We randomly selected images of 35 glaucoma subjects and 35 control subjects to compare diagnoses made by deep learning algorithms and clinicians. Her two glaucoma specialists (TO and SU) and her two pediatric ophthalmologists (TF and HS), who were blinded to the data collection procedure, determined whether each examination image corresponded to the glaucoma or control group. instructed to independently decide whether to They were informed that all images in the glaucoma group contained characteristic findings of childhood glaucoma, such as enlarged eyes and corneal opacification. The predictions of the deep learning models were determined by majority vote of the predictions of all five models.
