Data Preprocessing and Augmentation Improved Visual Field Prediction of Recurrent Neural Network with Multi-Central Datasets.

INTRODUCTION: The purpose of this study was to determine whether data preprocessing and augmentation could improve visual field (VF) prediction of recurrent neural network (RNN) with multi-central datasets. METHODS: This retrospective study collected data from five glaucoma services between June 2004 and January 2021. From an initial dataset of 331,691 VFs, we considered reliable VF tests with fixed intervals. Since the VF monitoring interval is very variable, we applied data augmentation using multiple sets of data for patients with more than eight VFs. We obtained 5,430 VFs from 463 patients and 13,747 VFs from 1,076 patients by setting the fixed test interval to 365 ± 60 days (D = 365) and 180 ± 60 days (D = 180), respectively. Five consecutive VFs were provided to the constructed RNN as input and the 6th VF was compared with the output of the RNN. The performance of the periodic RNN (D = 365) was compared to that of an aperiodic RNN. The performance of the RNN with 6 long- and short-term memory (LSTM) cells (D = 180) was compared with that of the RNN with 5-LSTM cells. To compare the prediction performance, the root mean square error (RMSE) and mean absolute error (MAE) of the total deviation value (TDV) were calculated as accuracy metrics. RESULTS: The performance of the periodic model (D = 365) improved significantly over aperiodic model. Overall prediction error (MAE) was 2.56 ± 0.46 dB versus 3.26 ± 0.41 dB (periodic vs. aperiodic) (p < 0.001). A higher perimetric frequency was better for predicting future VF. The overall prediction error (RMSE) was 3.15 ± 2.29 dB versus 3.42 ± 2.25 dB (D = 180 vs. D = 365). Increasing the number of input VFs improved the performance of VF prediction in D = 180 periodic model (3.15 ± 2.29 dB vs. 3.18 ± 2.34 dB, p < 0.001). The 6-LSTM in the D = 180 periodic model was more robust to worsening of VF reliability and disease severity. The prediction accuracy worsened as the false-negative rate increased and the mean deviation decreased. CONCLUSION: Data preprocessing with augmentation improved the VF prediction of the RNN model using multi-center datasets. The periodic RNN model predicted the future VF significantly better than the aperiodic RNN model.

Composition-Aware Image Steganography Through Adversarial Self-Generated Supervision.

Steganography is an important and prevailing information hiding tool to perform secret message transmission in an open environment. Existing steganography methods can mainly fall into two categories: predefined rule-based and data-driven methods. The former is susceptible to the statistical attack, while the latter adopts the deep convolution neural networks to promote security. However, deep learning-based methods suffer from perceptible artificial artifacts or deep steganalysis. In this article, we introduce a novel composition-aware image steganography (CAIS) to guarantee both visual security and resistance to deep steganalysis through the self-generated supervision. The key innovation is an adversarial composition estimation module, which has integrated the rule-based composition method and generative adversarial network to help synthesize steganographic images with more naturalness. We first perform a rule-based image blending method to obtain infinite synthetically data-label pairs. Then, we utilize an adversarial composition estimation branch to recognize the message feature pattern from the composite image based on these self-generated data-label pairs. Through the adversarial training, we force the steganography function to synthesize steganographic images, which can fool the composition estimation network. Thus, the proposed CAIS can achieve better information hiding and higher security to resist deep steganalysis. Furthermore, an effective global-and-part checking is designed to alleviate visual artifacts caused by hiding secret information. We conduct a comprehensive analysis of CAIS from various aspects (e.g., security and robustness) to verify the superior performance of the proposed method. Comprehensive experimental results on three large-scale widely used datasets have demonstrated the superior performance of our CAIS compared with several state-of-the-art approaches.

Visual field prediction using a deep bidirectional gated recurrent unit network model.

Although deep learning architecture has been used to process sequential data, only a few studies have explored the usefulness of deep learning algorithms to detect glaucoma progression. Here, we proposed a bidirectional gated recurrent unit (Bi-GRU) algorithm to predict visual field loss. In total, 5413 eyes from 3321 patients were included in the training set, whereas 1272 eyes from 1272 patients were included in the test set. Data from five consecutive visual field examinations were used as input; the sixth visual field examinations were compared with predictions by the Bi-GRU. The performance of Bi-GRU was compared with the performances of conventional linear regression (LR) and long short-term memory (LSTM) algorithms. Overall prediction error was significantly lower for Bi-GRU than for LR and LSTM algorithms. In pointwise prediction, Bi-GRU showed the lowest prediction error among the three models in most test locations. Furthermore, Bi-GRU was the least affected model in terms of worsening reliability indices and glaucoma severity. Accurate prediction of visual field loss using the Bi-GRU algorithm may facilitate decision-making regarding the treatment of patients with glaucoma.