Abstract: Abstract Background: The diagnosis of neuropsychological developmental disorders in children mainly depends on clinical manifestations at present, and there is still a lack of objective biological indicators. The eye characteristics of some children with neuropsychological developmental disorders may have important clinical value. Objective: To evaluate whether still eye photos can be used as a potential screening test for neuropsychological developmental disorders in children. Design: Case-control study. Methods: The Google image search engine was run with key words autism and child/autistic child/normal child/healthy child separately in order to select the sample images. Then the images were divided into two subsets, corresponding to population labeled with case and control respectively. Duplicated photos, cartoon photos, photos with facial deformity, photos without children's faces and photos with many children's faces hard to determine the research objects were deleted. Face tools identified and extracted the face out of each photograph. The resulting face images were scaled to 128 × 28 resolution (8-bit grayscale). Convolution neural network (CNN) was used in a machine installed with Keras, Scipy, Python imaging library (PIL). The binary cross entropy loss and rmsprop optimizer were used to train the model. The images were divided into training set, verification set and test set. The test score regarded as correct classification was more than 0.5 in case group and less than 0.5 or equal to 0.5 in the control group, and the training was stopped at 25 times. The grouping search results were regarded as the "gold standard", and the training group was regarded as the test standard to obtain the calculation accuracy and loss rate. Main outcome measures: Eye images generated from two sets of features. Results: Totally, 826 eye images were collected, 413 in the neuropsychological developmental disorders group and 413 in the control group. There were 736 (89.1%) in the training set, 44 (5.3%) in the validation set, and 46 (5.6%) in the test set. With increasing number of training sessions, the training and validation set accuracy rates showed an increasing trend while the training and validation loss rates gradually decreased. Test scores were 0.66±0.20 and 0.20±0.15, respectively in neuropsychological developmental disorder group and control group (t=9.03, P<0.001), and the number of correctly classified cases was 18/23 (78.3%) and 22/23 (95.6%). The mean accuracy of the two groups was 87.0%, the sensitivity was 78.3% (95% CI: 58.1%~90.3%), and the specificity was 95.6% (95% CI: 79.0%~99.2%). The receiver operating characteristic (ROC) curve of the trained model with deep learning showed the area under curve ROC (AUC) was 0.962. Eyes interval in the neuropsychological developmental disorder and control groups was generated by visualization of convolution layer of CNN, with the number of pixels being 90 and 70 in neuropsychological developmental disorder group and control group respectively. Conclusion: Supervised deep learning can obtain eye features of children with neuropsychological developmental disorders, which may be helpful to improve the early screening of neuropsychological developmental disorders by evaluating eye photos.

Key words: Deep learning, Eye photos, Children, Developmental disorders