Abstract
Purpose
To develop DeepISP, a deep learning model that predicts the comprehensive visual field (VF) information of the Humphrey visual field analyzer (HFA) based on rapid screening perimetry (Imo/TEMPO screening program [ISP]).
Design
A retrospective, cross-sectional, and longitudinal cohort database study.
Participants
One hundred eighty-seven actual ISPs from 112 patients who underwent both ISP and HFA 24-2 on the same day at the Jikei University School of Medicine Affiliated Hospital and 3470 synthesized ISPs from 883 patients who underwent VF measurements using HFA 24-2 and HFA 10-2 at 4 hospitals affiliated with Jikei University School of Medicine.
Methods
We developed 2 variants of multitask neural networks designed to predict both current VF parameters and VF progression parameters. We also evaluated the efficacy of data augmentation to synthesize ISP tests created by combining 20 points from HFA 24-2 and 8 points from HFA 10-2, with thresholding applied to these 28 points.
Main Outcome Measures
Mean absolute error for mean deviation (MD), pattern standard deviation (PSD), and visual field index (VFI). Mean F1 score for total deviation (TD) and pattern deviation (PD) probability plot classification. Area under the curve (AUC) for MD progression (MD slope <−1.0 decibel/year) and VFI progression (VFI slope <−1.8%/year).
Results
DeepISP could predict current VF status. Mean absolute errors for predicting MD, PSD, and VFI were 1.869 ± 0.114, 1.918 ± 0.082, and 5.146 ± 0.487, respectively. The mean F1 scores for pointwise classification of TD and PD probability plots were 0.761 ± 0.002 and 0.775 ± 0.002, respectively. The AUC for classifying glaucoma hemifield test was 0.920 ± 0.008. DeepISP was also capable of predicting VF progression, with AUCs of 0.828 ± 0.060 and 0.832 ± 0.062 for predicting MD and VFI progression, respectively.
Conclusions
We demonstrated ISP's versatility and capability in predicting comprehensive VF information, including current severity and progression risk. Our DeepISP serves as an efficient tool for screening and prioritizing patients with glaucoma for clinical intervention using only a single rapid ISP test.
Financial Disclosure(s)
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
Keywords: Glaucoma, Visual field, Perimetry, Screening, Deep learning
Glaucoma is a silently progressive visual field (VF) disorder that advances without early warning signs and is the leading cause of irreversible blindness worldwide.1,2 The global prevalence of glaucoma is increasing, with an estimated 112 million cases expected by 2040.3 Despite its significance, the only viable therapeutic approach to glaucoma is extending VF longevity by lowering intraocular pressure.4,5 Early detection and timely intervention are crucial for preserving the VF of patients with glaucoma.1,6,7 Therefore, comprehensive VF data, including the current severity, progression rate, and future state, are essential for managing glaucoma.
Structural and functional assessments are commonly used to evaluate both the presence and severity of glaucoma in clinical settings, such as OCT and VF tests. Moreover, multiple VF tests for over a year are needed to identify glaucoma progression.8,9
Although these clinical assessments are crucial for monitoring glaucoma, patients with mild glaucoma have limited access to these tests due to the lack of noticeable symptoms.1 Indeed, most patients with glaucoma remain undiagnosed globally.10, 11, 12 To address this, population-based screening may be a promising way to identify potential patients with glaucoma.13,14
In screening settings, structural tests with shorter examination times, such as fundus photographs and OCT, are widely used. Recent breakthroughs in deep learning technologies have enabled the prediction of current VF severity and defect patterns based on structural assessments.15, 16, 17, 18, 19 Furthermore, deep learning–based prediction of VF deterioration is now feasible by utilizing multiple structural examinations.20, 21, 22 However, predicting the VF progression rate or future state from a single structural assessment remains challenging. Given the temporal discrepancy between structural and functional changes in glaucoma,23 a function-based screening that directly evaluates the current VF potentially offers a more promising approach for assessing disease status and progression.
In this study, we propose DeepISP, a deep learning model that comprehensively predicts the current VF status and progression rate using a single functional test from the Imo/TEMPO screening program (ISP). The ISP is a rapid functional screening tool for glaucoma that employs either head-mounted or stationary perimetry (Imo/TEMPO),24, 25, 26 which assesses only 28 points of binary information (seen or not seen) and significantly reduces testing time for functional assessments in screening settings. Although ISP’s feasibility and performance for detecting glaucoma have been validated,24, 25, 26 its simplified test output limits the interpretation of detailed VF status. Beyond conventional screening for glaucoma prevalence, DeepISP aims to provide comprehensive VF information, including current status and progression rate, based on a single screening test, offering a more practical definition of glaucoma.
Furthermore, we demonstrated the effectiveness of data augmentation by synthesizing ISP tests from the widely utilized Humphrey visual field analyzer (HFA) tests. We also addressed external validity by training DeepISP on these synthesized ISP tests and evaluating its performance on an independent dataset of actual ISP tests. This approach can be replicated globally to address the scarcity of ISP test samples. Accordingly, we proposed a scalable model development strategy for DeepISP that facilitates high-resolution glaucoma screening.
Methods
Ethics
This study was approved by the ethics committee of Jikei University School of Medicine (approval no. 26-373[7879]). The study was designed in accordance with the tenets of the Declaration of Helsinki, and informed consent was obtained from all the patients.
Study Design and Participants
This study developed DeepISP, a multitask neural network model, to concurrently predict comprehensive VF information from HFA 24-2 tests using data from a rapid screening protocol, ISP. The ISP contains 28 points of binary information (seen or not seen) corresponding to the measurement points in HFA 24-2 or HFA 10-2. The predicted HFA 24-2 parameters included severity indicators (mean deviation [MD], pattern standard deviation [PSD], and visual field index [VFI]), VF defect patterns (total deviation [TD] maps and pattern deviation [PD] maps), and the glaucoma hemifield test (GHT) (Fig 1A). The model architecture was presented in Figure 1B and described in the following methods. A key challenge in developing DeepISP was the limited availability of ISP test data for training due to the novelty of this perimetric program. Thus, we employed data augmentation by synthesizing ISP tests derived from the widely available HFA test data (Fig 1C). This multicenter study included 187 actual ISP tests from 112 patients and 3470 synthesized ISP tests from 883 patients for the training models (Fig 1B).
Figure 1.
Overview of the DeepISP model and data process for model development. A, Design of the DeepISP model for predicting comprehensive VF information based on ISP. B, Patient inclusion flow diagram for actual and synthesized ISP set for the training models. C, Architecture of the multitask neural network model for predicting current VF parameters. D, Data processing for the synthesized ISP set derived from HFA 24-2 and HFA 10-2. E, Visualization of actual and synthesized ISP distributions with age using t-SNE. BCE = binary cross-entropy; CE = cross-entropy; GHT = glaucoma hemifield test; HFA = Humphrey visual field analyzer; ISP = Imo screening program; MD = mean deviation; PD = pattern deviation; PSD = pattern standard deviation; RMSE = root mean square error; TD = total deviation; t-SNE = t-distributed stochastic neighbor embedding; VF = visual field; VFI = visual field index.
Data Processing for Synthesized ISP
The ISP is a rapid perimetry test that uses a suprathreshold strategy with a stimulus intensity corresponding to <5% of normal sensitivity for that age at each test point. Each point can be stimulated up to 2 times if the initial response is missed; a response to any of these is judged as “seen,” otherwise “not seen.” Its 28 test points (binary values; seen or not seen) correspond to the measurement points in HFA 24-2 or HFA 10-2 (continuous values: light sensitivities). The synthesized ISP (28 points) was generated by combining pointwise sensitivities from HFA 24-2 (16 points; outside 10°) and HFA 10-2 (12 points; inside 10°), all conducted within a 6-month window at 4 hospitals affiliated with Jikei University School of Medicine (Fig 1C). These integrated sensitivities were binarized using age-specific thresholds for each participant, thereby reproducing 28 points of binary information to mimic the actual ISP tests. To validate the quality of the synthesized ISP for training DeepISP, we visualized the high-dimensional distributions of both the synthesized and actual ISPs using t-distributed stochastic neighbor embedding (t-SNE) (Fig 1D).
Datasets for Developing DeepISP
We retrospectively enrolled patients who underwent VF measurements using HFA 24-2 and HFA 10-2 at 4 hospitals affiliated with Jikei University School of Medicine between July 2007 and October 2023. Additionally, we included patients who underwent both ISP and HFA 24-2 on the same day at the Jikei University School of Medicine Affiliated Hospital between March 2018 and April 2022. The exclusion criteria were as follows: (1) patients who did not provide informed consent; (2) eyes with ocular findings other than glaucoma affecting the VFs in patients undergoing ISP; and (3) VF tests with low-reliability HFA 24-2 results. Following HFA 24-2 reliability criteria, all VF tests were required to meet the following standard: <15% false positives, <33% false negatives, and <20% fixation loss. The actual ISP, HFA 24-2, and HFA 10-2 tests that met these criteria were enrolled, and synthesized ISPs were generated based on these 2 HFA tests.
For patients who underwent multiple HFA 24-2 tests (more than 5), we calculated progression indicators (MD and VFI slopes) based on the HFA 24-2 test conducted on the same day and prior tests. For each patient, we identified time points at which ≥5 prior HFA 24-2 tests were available. At each such time point, we calculated the MD and VFI slopes using linear regression on all 24-2 tests preceding that time point, which are widely used methods for calculating these indicators.27,28 Only data with significant P values for the regression coefficients were included in this process. Furthermore, binary outcomes of MD and VFI progression were established using their respective cutoff values (−1.0 decibel/year for MD progression and −1.8%/year for VFI progression). The MD progression cutoff has been widely used, whereas the VFI progression cutoff was adopted from a recent study.22,27,28
Model Architecture Details of DeepISP
DeepISP involves predicting comprehensive VF information from ISP data. We developed 2 types of DeepISP models: model A for predicting current VF conditions and model B for retrieving VF progression information.
DeepISP model A is a multitask neural network designed to simultaneously predict current VF parameters such as MD, PSD, VFI, severity classifications of PD and TD maps, and the GHT class (Fig 1B).29 Model A incorporates 29 dimensions as input, combining 28 points from the ISP and age. The architecture includes a multilayer perceptron network with 2 hidden layers, each comprising 64 dimensions, and utilizes skip connections. The output layer integrates various task-specific layers to produce a 525-dimensional output. The GHT class is a binary classification (outer normal limits or not), whereas MD, PSD, and VFI are single scalar values, yielding 5 output dimensions. The TD and PD maps in the HFA 24-2 test classify severity into 5 classes at each of the 52 points (excluding the Mariott scotoma), accumulating a total output of 520 dimensions. The loss function was defined as follows: binary class entropy for the GHT class; root mean square error for MD, PSD, and VFI; and cross-entropy for the TD and PD maps. The overall network loss, defined as the weighted sum of these task-specific losses, was used to train the parameters via backpropagation. The weight coefficients were set based on the magnitude of each task-specific loss: 1 for MD, PSD, VFI, and GHT losses and 0.01 for TD and PD map losses. Other hyperparameters included a batch size of 32, a dropout rate of 0.1, a learning rate of 0.003, and a training duration of 10 epochs, with the Adam optimizer employed.
DeepISP model B is another multitask neural network focused on predicting progressive indices, such as the slopes of MD and VFI, along with the binary outcomes of MD and VFI progression (as shown in Fig 2A). Model B constitutes a multilayer perceptron network with the same 29-dimensional input as model A and 2 hidden layers, each featuring 16 dimensions and utilizing skip connections. The output layer comprises 4 task-specific layers, yielding 6 output dimensions. The MD and VFI slopes were single scalar values, and the MD and VFI progressions were binary classifications. The loss functions were designated using root mean square error for the slopes and binary class entropy for the progression tasks. The training parameters leveraged the aggregate network loss, with all weight coefficients standardized at 1. Additional hyperparameters included a batch size of 32, a dropout rate of 0.1, a learning rate of 0.003, and a training duration of 10 epochs, also utilizing the Adam optimizer.
Figure 2.
DeepISP predicts the current VF progression. A, Patients inclusion flow diagram for actual and synthesized ISP sets for predicting current VF progression. B, Visualization of actual and synthesized ISP distributions with age features for predicting current VF progression using t-SNE. C, Visualization of ISP data distribution with the age feature using t-SNE, with representations colored according to the current VF progression indices (MD and VFI slopes). D, Average predictions for MD and VFI slopes based on the synthetic-based model trained with the synthesized ISP across 100-trial randomization. E, AUC performance for predicting MD and VFI progression based on a synthetic model trained with synthesized ISP using 100-trial seed randomization. F, Feature importance of ISP measurement points emphasized when predicting MD and VFI slopes. AUC = area under receiver operating characteristic curve; ISP = Imo screening program; MD = mean deviation; t-SNE = t-distributed stochastic neighbor embedding; VF = visual field; VFI = visual field index.
Evaluation of Model Performance of DeepISP
In evaluating DeepISP model A for predicting current VF information, 3 variants were created based on a combination of training and evaluation data. Model A1 was trained solely on actual ISP data and evaluated on a separate set of unused actual ISPs. Model A2 was trained exclusively on synthesized ISPs and assessed for all actual ISPs. Model A3 was initially trained on synthesized ISPs and then further trained on actual ISPs, which was the evaluation domain, and evaluated on an unused actual ISP. Leave-one-out cross-validation was applied to evaluate models A1 and A3. DeepISP model B, which focuses on VF progression information, was trained only on synthesized ISPs and assessed on all actual ISPs, given the superior performance of the synthetic-based model in predicting current VF information. Model performance was assessed using mean absolute error (MAE), area under the curve (AUC), and macro F1-score for continuous value predictions, binary classifications, and multiclass classifications, respectively. All models were trained and evaluated using 100-trial seed randomization.
Feature Importance for Predicting VF Progression Information
Feature importance for each ISP measurement point was calculated to predict the MD and VFI slopes (as shown in Fig 2F). For each ISP measurement feature, model performance (MAE for predicting MD and VFI slopes) was evaluated by altering the feature values using “1 - original value (seen = 1 or not = 0)” across the test dataset, which is one of the applications of permutation importance. The importance of each feature was determined by measuring the absolute difference in model performance before and after feature value alternation. This process was repeated for each feature, and the feature importance scores were averaged across multiple models using 100-trial seed randomization. The mean feature importance scores were scaled to sum to 100% for interpretability.
Archetypal Analysis of ISP Tests
Archetypal analysis is a matrix factorization method to approximate data points with a convex hull.30, 31, 32, 33 The factorization yields extreme points of the convex hull, called “archetypes (ATs).” A total of 3470 synthesized ISPs (Fig 1B) were visualized through archetypal analysis, and the optimal number of ATs was determined using the elbow method, which identifies the point of diminishing returns in reconstruction error reduction as the number of ATs increases. The synthesized ISPs were categorized into 2 groups based on the composition of each AT: one group had a higher composition than the random assignment (>1/7), whereas the other had a lower composition (<1/7). We demonstrated the distribution of MD values in the group with a higher composition of each AT.
Furthermore, we investigated the association of each AT with the future MD declines (+1-year ΔMD), calculated from data derived from more than 5 serial HFA 24-2 tests. We identified time points where 5 or more subsequent HFA 24-2 tests were available. At each time point, we computed the future MD declines using linear regression over exactly 5 subsequent 24-2 tests. Only time points with statistically significant slopes were retained for analysis. This approach considers only the 5 most recent future data points and is adjusted to represent the average annual change. The median future MD values among the groups with higher or lower compositions of each AT were compared using a 2-sided Mann–Whitney U test.
Results
Characteristics of the Study Participants
This multicenter study included 187 actual ISP tests from 112 patients and 3470 synthesized ISP tests from 883 patients for the training models (Fig 1B). Table 1 and Figure S1 (available at www.ophthalmologyscience.org) summarize the characteristics of the participants in both the synthesized and actual ISP groups. The mean ± standard deviation for the synthesized and actual ISP groups, respectively, were as follows: MD (−9.10 ± 8.41 and −3.86 ± 4.82), PSD (8.22 ± 4.65 and 5.45 ± 4.31), and VFI (73.3 ± 25.3 and 89.5 ± 13.1). The synthesized ISP group showed greater variability in severity than the ISP group, representing data diversity. As demonstrated by t-SNE in high-dimensional feature space (Fig 1D), the synthesized ISP not only covered the distribution of the actual ISP samples but also ensured abundant samples, even in regions where the actual ISP samples were sparse.
Table 1.
Characteristics of ISP Data for Predicting Current VF Details
| Synthesized ISP | Actual ISP | P Value | |
|---|---|---|---|
| Number of Tests | 3470 | 187 | |
| Age (yrs) | 60.2 ± 12.9 | 61.4 ± 12.9 | 0.263 |
| MD (dB) | −9.10 ± 8.41 | −3.86 ± 4.82 | <0.001 |
| PSD (dB) | 8.22 ± 4.65 | 5.45 ± 4.31 | <0.001 |
| VFI (%) | 73.3 ± 25.3 | 89.5 ± 13.1 | <0.001 |
| GHT, N (%) | 2798 (80.6%) | 126 (67.4%) | <0.001 |
dB = decibel; GHT = glaucoma hemifield test; ISP = Imo screening program; MD = mean deviation; PSD = pattern standard deviation; VF = visual field; VFI = visual field index.
Statistical significance is indicated (P < 0.05, 2-sided Mann–Whitney U test). Bold values indicate a statistically significant difference between 2 groups.
DeepISP Reproduces Current VF Details
The ISP tests demonstrated distinguishable features for identifying current VF severity (MD, PSD, and VFI), as visualized using t-SNE in high-dimensional feature space (Fig 3A). This finding validates the feasibility of predicting glaucoma severity using synthesized or actual ISP data.
Figure 3.
DeepISP reproduces the current VF details. A, Visualization of ISP data distribution with age features using t-SNE, with representations colored according to current VF severity indices (MD, PSD, VFI). B, Violin plots to compare the predictive performance for TD and PD maps, MD, PSD, and VFI based on 3 models using 100-trial seed randomization: Model A1 (trained with actual ISP), Model A2 (trained with synthesized ISP), and Model A3 (trained with both datasets). C, Predictive performance for TD and PD maps based on Model A2 (synthetics-based model trained with synthesized ISP) across 100-trial randomization. D, Average predictions for MD, PSD, and VFI based on Model A2 (a synthetics-based model trained with synthesized ISP) across 100-trial randomization. E, Comparison of AUC performance for predicting GHT based on the 3 models using 100-trial seed randomization. F, Ground truth and model predictions for representative HFA 24-2 test sample based on ISP using the synthetics-based model. AUC = area under receiver operating characteristic curve; GHT = glaucoma hemifield test; HFA = Humphrey visual field analyzer; ISP = Imo screening program; MAE = mean absolute error; MD = mean deviation; O.N.L. = outer normal limits; PD = pattern deviation; PSD = pattern standard deviation; TD = total deviation; t-SNE = t-distributed stochastic neighbor embedding; VF = visual field; VFI = visual field index.
When comparing the predictive performance of the 3 models (A1, A2, and A3), model A2 demonstrated the highest predictive performance in most tasks, followed by models A3 and A1 (Table 2, Fig 3B, E). Model A2 functions as a synthetic-based model that does not rely on the actual ISP. Nonetheless, this model can predict VF severity (MD, PSD, and VFI), glaucoma likelihood (GHT), and VF defect patterns (TD and PD maps). For instance, the MAEs for predicting MD, PSD, and VFI were 1.869 ± 0.114, 1.918 ± 0.082, and 5.146 ± 0.487, respectively. The mean F1 scores for pointwise classification of the TD and PD maps were 0.761 ± 0.002 and 0.775 ± 0.002, respectively. The AUC for classifying GHT was 0.920 ± 0.008.
Table 2.
Performance of DeepISP to Predict Current VF Details
| Model A1 | Model A2 | Model A3 | P Value (2 vs. 1) | P Value (2 vs. 3) | |
|---|---|---|---|---|---|
| TD map (F1 score) | 0.729 ± 0.003 | 0.761 ± 0.002 | 0.759 ± 0.002 | <0.001 | <0.001 |
| PD map (F1 score) | 0.749 ± 0.003 | 0.775 ± 0.002 | 0.776 ± 0.002 | <0.001 | <0.001 |
| MD (MAE) | 2.629 ± 0.132 | 1.869 ± 0.114 | 1.974 ± 0.052 | <0.001 | <0.001 |
| PSD (MAE) | 2.267 ± 0.099 | 1.918 ± 0.082 | 2.037 ± 0.048 | <0.001 | <0.001 |
| VFI (MAE) | 7.431 ± 0.399 | 5.146 ± 0.487 | 5.272 ± 0.204 | <0.001 | <0.001 |
| GHT (AUROC) | 0.902 ± 0.009 | 0.920 ± 0.008 | 0.909 ± 0.007 | <0.001 | <0.001 |
AUC = area under receiver operating characteristic curve; GHT = glaucoma hemifield test; ISP = Imo screening program; MAE = mean absolute error; MD = mean deviation; PD = pattern deviation; PSD = pattern standard deviation; TD = total deviation; VF = visual field; VFI = visual field index.
Statistical significance is indicated (P < 0.05/12, 2-sided Mann–Whitney U test). Bold values indicate a statistically significant difference between the 2 groups.
Figure 3C–E illustrate the predictive performance of the synthetic-based model with the highest accuracy. Pointwise predictions of TD and PD maps exhibited relatively higher accuracy in the central and superior VF regions. Scatter plots of MD and VFI indicated general agreement between ground truth and predicted values. Glaucoma hemifield test predictions achieved an AUC exceeding 0.9 for all 3 models. Additionally, Fig 3F presents ground truth and predicted results for a representative test sample using the synthetic-based model. For the PD map, the model provided accurate pointwise predictions for both normal and most severe classes. In summary, ISP has significant potential for predicting current VF information from HFA 24-2 data.
DeepISP Predicts the Current VF Progression
A multitask model B was developed to predict VF progressive indicators such as the MD and VFI slopes and MD and VFI progression (Fig 2A). To develop model B, 731 synthesized ISPs from 214 patients and 60 actual ISPs from 36 patients were included to predict these VF progressive indicators (Fig 2A).
Table 3 and Figure S2 (available at www.ophthalmologyscience.org) summarize the characteristics of the participants with progressive information in both groups. The mean ± standard deviation values for the synthesized and actual ISP groups, respectively, were as follows: MD (−9.12 ± 7.37 and −3.95 ± 4.75), PSD (8.88 ± 4.57 and 5.48 ± 3.94), VFI (73.9 ± 22.2 and 89.3 ± 13.1), MD slope (−0.61 ± 0.82 and −0.37 ± 0.64), and VFI slope (−1.7 ± 2.7 and −0.8 ± 1.4).
Table 3.
Characteristics of ISP Data for Predicting Current VF Progressions
| Synthesized ISP | Actual ISP | P Value | |
|---|---|---|---|
| Number of tests | 731 | 60 | |
| Age (yrs) | 60.6 ± 12.4 | 61.1 ± 12.8 | 0.834 |
| MD (dB) | −9.12 ± 7.37 | −3.95 ± 4.75 | <0.001 |
| MD slope (dB/yr) | −0.61 ± 0.82 | −0.37 ± 0.64 | 0.012 |
| PSD (dB) | 8.88 ± 4.57 | 5.48 ± 3.94 | <0.001 |
| VFI (%) | 73.9 ± 22.2 | 89.3 ± 13.1 | <0.001 |
| VFI slope (%/yr) | −1.7 ± 2.7 | −0.8 ± 1.4 | 0.004 |
| GHT, N (%) | 618 (84.5%) | 43 (71.7%) | 0.016 |
dB = decibel; GHT = glaucoma hemifield test; ISP = Imo screening program; MD = mean deviation; PSD = pattern standard deviation; VF = visual field; VFI = visual field index.
Statistical significance is indicated (P < 0.05, 2-sided Mann–Whitney U test). Bold values indicate a statistically significant difference between the 2 groups.
Feature visualization by t-SNE of these ISPs revealed that the synthesized ISP covered the distribution of actual ISP samples while exhibiting a more diverse distribution (Fig 2B). Additionally, both the synthesized and actual ISPs exhibited distinct features based on the VF progression information in high-dimensional feature space (Fig 2C). This finding validates the feasibility of predicting VF progression risk using the synthesized ISP.
Given the high performance of the synthetic-based model in predicting current VF information, we developed a synthetic-based model to predict VF progression information using only the synthesized ISPs. Table 4, Figure 2D,E present the details of the predictive performance of the synthetic-based model. For instance, the MAEs for predicting MD and VFI slopes were 0.422 ± 0.019 and 0.893 ± 0.059, respectively. Scatter plots of these slopes indicate a rough agreement between ground truth and predicted values (Fig 2D). Moreover, we assessed the AUC performance in screening progressive cases based on thresholding of the progression rate; the AUCs for predicting the binary outcomes of MD and VFI progression were 0.828 ± 0.060 and 0.832 ± 0.062, respectively (Fig 2E).
Table 4.
Performance of DeepISP to Predict Current VF Progressions
| Model B | |
|---|---|
| MD slope (MAE) | 0.422 ± 0.019 |
| VFI slope (MAE) | 0.893 ± 0.059 |
| MD progression (AUC) | 0.828 ± 0.060 |
| VFI progression (AUC) | 0.832 ± 0.062 |
AUC = area under receiver operating characteristic curve; ISP = Imo screening program; MAE = mean absolute error; MD = mean deviation; VF = visual field; VFI = visual field index.
Furthermore, we enhanced the interpretability of DeepISP by presenting the pointwise feature importance of the ISP measurement points for predicting MD and VFI slopes (Fig 2F). The inferonasal paracentral point consistently emerged as critically important.
Associations between Latent Archetypal Patterns of ISP and Future VF State
A detailed assessment of the future VF state allows for a more precise patient risk stratification than progression rates alone. To approximate the future state, we calculated the future MD declines (+1-year ΔMD) using data from more than 5 serial HFA 24-2 tests. In this analysis, we included 774 synthesized ISPs from 264 patients. Table 5 summarizes their characteristics.
Table 5.
Characteristics of ISP Data for Archetypal Analyses Using Synthesized ISP
| Synthesized ISP | |
|---|---|
| Number of tests | 774 |
| Age (yrs) | 58.3 ± 12.3 |
| MD (dB) | −7.83 ± 6.91 |
| +1 yr ΔMD | −0.48 ± 0.78 |
| PSD (dB) | 8.63 ± 4.74 |
| VFI (%) | 77.6 ± 19.7 |
| +1 yr ΔVFI | −1.4 ± 2.4 |
| GHT, N (%) | 647 (83.6%) |
dB = decibel; GHT = glaucoma hemifield test; ISP = Imo screening program; MD = mean deviation; PSD = pattern standard deviation; VFI = visual field index.
Future MD declines showed low concordance with the MD slopes and can be considered distinct metrics (Fig. S3A, available at www.ophthalmologyscience.org). No distinguishable features of future MD decline were identified in the high-dimensional space visualized using t-SNE (Fig S3B). Thus, rather than predicting future MD decline, we focused on assessing the association between future MD decline and the local distribution of latent ISP representations.
To achieve this, we conducted an archetypal analysis to generate extreme patterns (ATs) based on high-dimensional data features, representing individual data as mixtures of these ATs (Fig 4A).30, 31, 32, 33 Synthesized ISPs were divided into higher and lower AT groups based on AT composition. Comparisons of the mean MD values within the higher AT group revealed variability in VF severity across ATs, indicating diversity in their features (Fig 4B). To examine the association between future MD decline and each AT, we analyzed the differences in future MD decline between the higher and lower groups for each AT (Fig 4C). The inferior altitudinal (AT6) and superior paracentral (AT1) VF defect patterns demonstrated a higher risk of progression. In contrast, the normal AT (AT5) was associated with a lower risk of progression.
Figure 4.
Associations between latent archetypal patterns of ISP and future VF state. A, Seven archetypal patterns (AT1–AT7) of the synthesized ISP were identified by archetypal analysis. B, Simplex projection of archetypal analysis colored by MD values (left) and distribution of MD values for each archetype with a composition greater than the random expectation (> 1/7) (right). C, Box plots comparing future MD declines (+1-year ΔMD) between low (composition < 1/7) and high (composition > 1/7) composition groups for each archetype. Statistical significance was indicated (P < 0.05/7, 2-sided Mann–Whitney U test). ISP = Imo screening program; MD = mean deviation; VF = visual field.
Discussion
In this study, we developed a DeepISP model to predict detailed VF information in the HFA 24-2 test based on the simple binary test information from ISP. During model development, we investigated the effectiveness of data augmentation by synthesizing ISP data derived from the widely used HFA 24-2 and HFA 10-2 tests. Our findings demonstrated that the synthetic-based model trained solely on the synthesized ISP achieved the best performance. This approach enables model development at various facilities worldwide without relying on scarce ISP tests, providing a scalable model development strategy. To screen and prioritize patients requiring more urgent intervention, we stratified patient-specific risks based on current progression rates and future MD decline. Additionally, we identified key VF test points and defect patterns critical for assessing individualized risks. Despite the simplicity of the ISP test information, these results demonstrate the potential of DeepISP to comprehensively assess VF information, including current VF details and progression, from a single ISP test.
Our model could reproduce the current VF details in HFA 24-2 using the ISP test. Notably, the predictions for current VF severity indices, such as MD and VFI, were particularly reliable (Fig 3D). The predictions for the GHT class achieved a high AUC score across all models (Fig 3E), indicating its ability to detect glaucoma. Furthermore, DeepISP provided pointwise predictions for the TD and PD map classes (Fig 3C), with stronger agreement in the normal and most severe cases (Fig 3F). However, our model introduced false positives and negatives at certain test points (Fig 3F). False positives likely stemmed from overestimating central VF defects in paracentral areas not captured by HFA 24-2, potentially providing more granular information than the ground truth data (HFA 24-2). Conversely, false negatives were likely observed in cases where VF defects occurred in peripheral regions beyond the ISP test coverage.
We trained 3 models to predict current VF information: model A1, using the actual ISP; model A2, with the synthesized ISP; and model A3, combining both datasets. Among these, model A2, which was externally validated using independent-domain datasets for training and evaluation, exhibited the best performance, followed by models A3 and A1 (Table 2, Fig 3B, E). The superior performance of model A2 over model A1 can be attributed to the difference in sample sizes. Owing to the abundance of test samples from the HFA, the synthesized dataset included a broader range of VF severities, which likely contributed to the improved performance and enhanced model representation. However, model A2’s superiority over model A3 cannot be attributed solely to sample size because model A3 had a larger sample size. One possible explanation is the higher quality of the synthesized ISP data compared with the actual ISP, which may have caused negative transfer when using the actual ISPs.34 The thorough threshold exploration of the HFA measurement rules likely improved the data quality of the synthesized ISP derived from the HFA tests. Furthermore, training with the synthesized ISP leverages the robust correspondence between the inputs (HFA 24-2) and outputs (synthesized ISP derived from HFA 24-2), potentially enhancing the stability of the model parameter training.
We also predicted VF progression using multiple HFA 24-2 tests based on a single ISP test. Integrating predictions for current VF severity and progression enables precise patient screening and prioritization for clinical interventions. Although previous studies have attempted to predict VF progression using multiple structural tests,20, 21, 22 this study aimed to obtain comprehensive VF information from only a single screening test. The effectiveness of synthetic-based learning in predicting VF progression further supports the versatility of data augmentation using synthesized ISPs across various tasks. Predictions of MD and VFI progression showed strong performance, with AUCs of 0.828 and 0.832, respectively (Table 4 and Fig 2E), underscoring their clinical applicability. Additionally, we demonstrated the potential importance of the inferonasal paracentral point in ISP tests for determining VF progression (Fig 2F). Two plausible explanations exist for the association between inferior VF defects and progression. First, most cases with inferior VF defects often co-occur with superior defects,35,36 leading to multiple VF defects that may have increased the risk of further progression due to the continuous expansion of existing glaucomatous VF defects.37 Second, inferior VF defects may reflect atypical characteristics, such as reduced retinal blood flow. Although superior VF defects are more common in glaucoma,38,39 inferior VF defects are linked to reduced blood flow in the superior retina, which is associated with a higher risk of progression.38 Therefore, the inferonasal paracentral point in ISP tests may be a critical marker for evaluating glaucoma progression.
We focused on predicting future MD decline as an approximate indicator of VF status 1 year later, which is crucial for determining optimal short-term treatment strategies. However, the future MD decline was poorly represented within the ISP feature space (Fig S3B), suggesting challenges in directly predicting this indicator. Instead, we identified 3 ATs of latent ISP representations as factors associated with future MD decline (Fig 4A, C). First, the inferior altitudinal VF defect pattern (AT6) was related to future VF deterioration, aligning with the findings of previous studies on HFA 24-2 VF patterns.31,33 Notably, this pattern includes the inferonasal paracentral point, which supports our previous findings regarding current VF progres sion. Second, the superior paracentral VF defect pattern (AT1) aligns with prior research on HFA 24-2 VF patterns,33 and involves typical VF defect areas observed in mild glaucoma, where early detection is critical for timely intervention.40 Finally, the reduced normal AT (AT5) was also associated with future VF deterioration. These findings on VF progression and future VF states have the potential to enhance clinical decision-making for personalized assessments and interventions.
Our study had some limitations. First, the ISP sample size in the evaluation dataset was limited. We addressed this through data augmentation and synthetic-based model development by creating a synthesized ISP. Second, the DeepISP was trained and evaluated in a Japanese population. Although training on synthesized ISP contributed to improved performance on actual ISP in our population, this finding may not necessarily generalize to other populations due to differences in average age and disease prevalence. To enable broader adoption, external validation across diverse demographic datasets from training to evaluation is required. Additionally, empirical studies in screening settings are warranted. We believe that our proposed strategy for synthesizing ISP tests will support further validation efforts. Third, accurately predicting future VF states remains challenging. However, we gained insights by stratifying the risks of future MD decline based on latent representation patterns. Lastly, as a retrospective study, we could not account for unidentified confounders, such as intraocular pressure and myopia status, which are major factors for glaucoma. These factors are relevant in screening settings and incorporating them could further enhance DeepISP.
In conclusion, this study demonstrated ISP's versatility and capability in predicting comprehensive VF information, including current severity, progression risk, and partial future states. By leveraging ISP, DeepISP serves as an efficient tool for screening and prioritizing patients with glaucoma for clinical intervention using only a single rapid ISP test. Furthermore, the superior predictive performance of the synthetic-based model highlights that globally utilized HFA could enhance ISP’s screening potential to evaluate glaucoma. Our reproducible model development strategy, with its scalability across different domains, would facilitate the worldwide adoption of ISPs. The global dissemination of high-resolution screening with DeepISP could pave the way for a new era of personalized glaucoma management.
Data Availability
Not all clinical data utilized for training or evaluating DeepISP are publicly accessible. For noncommercial purposes, requests to access clinical data about visual tests and related clinical information should be directed to T.N. or E.N. The development codes for DeepISP and synthesizing ISP are available on our GitHub repository (https://github.com/keimy1007/DeepISP).
Acknowledgments
The authors want to thank Editage (www.editage.com) for English language editing.
Manuscript no. XOPS-D-24-00501.
Footnotes
Supplemental material available atwww.ophthalmologyscience.org.
Disclosures:
All authors have completed and submitted the ICMJE disclosures form.
The authors made the following disclosures:
A.I.: Honoraria – Carl Zeiss Meditec, CREWT Medical Systems, Heidelberg Engineering, Santen Pharmaceutical Co, Ltd, Senju Pharmaceutical Co, Ltd, Otsuka Pharmaceutical Co, Ltd (lectures).
T. Noro: Consultant – Santen Pharmaceutical Co., Ltd.; Honoraria – Carl Zeiss Meditec, Inc., Santen Pharmaceutical Co., Kowa Company, Ltd., TOPCON CORPORATION.
T. Nakano: Grants – Kyowa Medical, HOYA CORPORATION, Senju Pharmaceutical, IOL MEDICAL, Japan Airlines, Sanofi, KURIBARA MEDICAL INSTRUMENTS, Otsuka Pharmaceutical, Santen Pharmaceutical, All Nippon Airway, Carl Zeiss.
This study was supported by the Japan Society for the Promotion of Science (KAKENHI, Grant Number JP24K19797). These funding organizations had no role in the design or conduct of this research.
HUMAN SUBJECTS: Human subjects were included in this study. This study was approved by the ethics committee of Jikei University School of Medicine (approval no. 26-373[7879]). The study was designed in accordance with the tenets of the Declaration of Helsinki, and informed consent was obtained from all the patients.
No animal subjects were used in this study.
Author Contributions:
Conception and design: Sano, Nishijima, Sumi
Data collection: Sano, Nishijima, Sumi, Noro, Ogawa, Igari, Iwase, Nakano
Analysis and interpretation: Sano, Nishijima, Sumi, Noro, Igari
Obtained funding: N/A
Overall responsibility: Sano, Nishijima, Sumi
Contributor Information
Euido Nishijima, Email: e.nishijima@jikei.ac.jp.
Shunsuke Sumi, Email: shunsukesumi@iqb.u-tokyo.ac.jp.
Tadashi Nakano, Email: tnakano@jikei.ac.jp.
Supplementary Data
References
- 1.Weinreb R.N., Aung T., Medeiros F.A. The pathophysiology and treatment of glaucoma: a review. JAMA. 2014;311:1901–1911. doi: 10.1001/jama.2014.3192. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Beykin G., Norcia A.M., Srinivasan V.J., et al. Discovery and clinical translation of novel glaucoma biomarkers. Prog Retin Eye Res. 2021;80:100875. doi: 10.1016/j.preteyeres.2020.100875. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Tham Y.C., Li X., Wong T.Y., et al. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology. 2014;121:2081–2090. doi: 10.1016/j.ophtha.2014.05.013. [DOI] [PubMed] [Google Scholar]
- 4.Musch D.C., Gillespie B.W., Niziol L.M., et al. Intraocular pressure control and long-term visual field loss in the Collaborative Initial Glaucoma Treatment Study. Ophthalmology. 2011;118:1766–1773. doi: 10.1016/j.ophtha.2011.01.047. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.De Moraes C.G., Liebmann J.M., Levin L.A. Detection and measurement of clinically meaningful visual field progression in clinical trials for glaucoma. Prog Retin Eye Res. 2017;56:107–147. doi: 10.1016/j.preteyeres.2016.10.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Chen P.P. Blindness in patients with treated open-angle glaucoma. Ophthalmology. 2003;110:726–733. doi: 10.1016/S0161-6420(02)01974-7. [DOI] [PubMed] [Google Scholar]
- 7.Saunders L.J., Russell R.A., Kirwan J.F., et al. Examining visual field loss in patients in glaucoma clinics during their predicted remaining lifetime. Invest Ophthalmol Vis Sci. 2014;55:102–109. doi: 10.1167/iovs.13-13006. [DOI] [PubMed] [Google Scholar]
- 8.Chauhan B.C., Garway-Heath D.F., Goni F.J., et al. Practical recommendations for measuring rates of visual field change in glaucoma. Br J Ophthalmol. 2008;92:569–573. doi: 10.1136/bjo.2007.135012. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Crabb D.P., Garway-Heath D.F. Intervals between visual field tests when monitoring the glaucomatous patient: wait-and-see approach. Invest Ophthalmol Vis Sci. 2012;53:2770–2776. doi: 10.1167/iovs.12-9476. [DOI] [PubMed] [Google Scholar]
- 10.Chua J., Baskaran M., Ong P.G., et al. Prevalence, risk factors, and visual features of undiagnosed glaucoma: the Singapore epidemiology of eye diseases study. JAMA Ophthalmol. 2015;133:938–946. doi: 10.1001/jamaophthalmol.2015.1478. [DOI] [PubMed] [Google Scholar]
- 11.Shaikh Y., Yu F., Coleman A.L. Burden of undetected and untreated glaucoma in the United States. Am J Ophthalmol. 2014;158:1121–1129.e1. doi: 10.1016/j.ajo.2014.08.023. [DOI] [PubMed] [Google Scholar]
- 12.Iwase A., Suzuki Y., Araie M., et al. The prevalence of primary open-angle glaucoma in Japanese: the Tajimi Study. Ophthalmology. 2004;111:1641–1648. doi: 10.1016/j.ophtha.2004.03.029. [DOI] [PubMed] [Google Scholar]
- 13.Tang J., Liang Y., O'Neill C., et al. Cost-effectiveness and cost-utility of population-based glaucoma screening in China: a decision-analytic Markov model. Lancet Glob Health. 2019;7:e968–e978. doi: 10.1016/S2214-109X(19)30201-3. [DOI] [PubMed] [Google Scholar]
- 14.Aspberg J., Heijl A., Bengtsson B. Screening for open-angle glaucoma and its effect on blindness. Am J Ophthalmol. 2021;228:106–116. doi: 10.1016/j.ajo.2021.03.030. [DOI] [PubMed] [Google Scholar]
- 15.Chen Z., Ishikawa H., Wang Y., et al. Deep-learning-based group pointwise spatial mapping of structure to function in glaucoma. Ophthalmol Sci. 2024;4:100523. doi: 10.1016/j.xops.2024.100523. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Lazaridis G., Montesano G., Afgeh S.S., et al. Predicting visual fields from optical coherence tomography via an ensemble of deep representation learners. Am J Ophthalmol. 2022;238:52–65. doi: 10.1016/j.ajo.2021.12.020. [DOI] [PubMed] [Google Scholar]
- 17.Kihara Y., Montesano G., Chen A., et al. Policy-driven, multimodal deep learning for predicting visual fields from the optic disc and OCT imaging. Ophthalmology. 2022;129:781–791. doi: 10.1016/j.ophtha.2022.02.017. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Yu H.H., Maetschke S.R., Antony B.J., et al. Estimating global visual field indices in glaucoma by combining macula and optic disc OCT scans using 3-dimensional convolutional neural networks. Ophthalmol Glaucoma. 2021;4:102–112. doi: 10.1016/j.ogla.2020.07.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Christopher M., Bowd C., Belghith A., et al. Deep learning approaches predict glaucomatous visual field damage from OCT optic nerve head en face images and retinal nerve fiber layer thickness maps. Ophthalmology. 2020;127:346–356. doi: 10.1016/j.ophtha.2019.09.036. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Nouri-Mahdavi K., Mohammadzadeh V., Rabiolo A., et al. Prediction of visual field progression from OCT structural Measures in moderate to advanced glaucoma. Am J Ophthalmol. 2021;226:172–181. doi: 10.1016/j.ajo.2021.01.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Mohammadzadeh V., Wu S., Davis T., et al. Prediction of visual field progression with serial optic disc photographs using deep learning. Br J Ophthalmol. 2023;108:1107–1113. doi: 10.1136/bjo-2023-324277. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Hou K., Bradley C., Herbert P., et al. Predicting visual field worsening with longitudinal optical coherence tomography data using a gated transformer network (OPHTHA-D-22-01964)s. Ophthalmology. 2023;130:854–862. doi: 10.1016/j.ophtha.2023.03.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Moghimi S., Bowd C., Zangwill L.M., et al. Measurement floors and dynamic ranges of OCT and OCT angiography in glaucoma. Ophthalmology. 2019;126:980–988. doi: 10.1016/j.ophtha.2019.03.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Arai K., Nishijima E., Ogawa S., et al. A novel visual field screening program for glaucoma with a head-mounted perimeter. J Glaucoma. 2023;32:520–525. doi: 10.1097/IJG.0000000000002199. [DOI] [PubMed] [Google Scholar]
- 25.Nishijima E., Hosaka D., Ogawa S., et al. Comparison of glaucoma detection performance of binocular perimetry screening program using imo perimetry with frequency doubling technology. Transl Vis Sci Technol. 2024;13:9. doi: 10.1167/tvst.13.8.9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Nishijima E., Itoh Y., Noro T., et al. Correlation of binocular perimetry screening using imo, a portable head mounted perimeter with 10-2 standard automated perimetry for early glaucoma with central visual field defects. J Glaucoma. 2024;34:89–94. doi: 10.1097/IJG.0000000000002492. [DOI] [PubMed] [Google Scholar]
- 27.Saeedi O.J., Elze T., D'Acunto L., et al. Agreement and predictors of discordance of 6 visual field progression algorithms. Ophthalmology. 2019;126:822–828. doi: 10.1016/j.ophtha.2019.01.029. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Rabiolo A., Morales E., Mohamed L., et al. Comparison of methods to detect and measure glaucomatous visual field progression. Transl Vis Sci Technol. 2019;8:2. doi: 10.1167/tvst.8.5.2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Ruder S. An overview of multi-task learning in deep neural networks. arXiv. 2017 doi: 10.48550/arXiv.1706.05098. [DOI] [Google Scholar]
- 30.Cutler A., Breiman L. Archetypal analysis. Technometrics. 1994;36:338–347. [Google Scholar]
- 31.Wang M., Shen L.Q., Pasquale L.R., et al. An artificial intelligence approach to detect visual field progression in glaucoma based on spatial pattern analysis. Invest Ophthalmol Vis Sci. 2019;60:365–375. doi: 10.1167/iovs.18-25568. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Doshi H., Solli E., Elze T., et al. Unsupervised machine learning shows change in visual field loss in the idiopathic intracranial hypertension treatment trial. Ophthalmology. 2022;129:903–911. doi: 10.1016/j.ophtha.2022.03.027. [DOI] [PubMed] [Google Scholar]
- 33.Yousefi S., Pasquale L.R., Boland M.V., Johnson C.A. Machine-identified patterns of visual field loss and an association with rapid progression in the ocular hypertension treatment study. Ophthalmology. 2022;129:1402–1411. doi: 10.1016/j.ophtha.2022.07.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Shao L., Zhu F., Li X. Transfer learning for visual categorization: a survey. IEEE Trans Neural Netw Learn Syst. 2015;26:1019–1034. doi: 10.1109/TNNLS.2014.2330900. [DOI] [PubMed] [Google Scholar]
- 35.Germano R.A.S., Germano C.S., Susanna F.N., Susanna R. Patterns of visual field loss in early, moderate, and severe stages of open angle glaucoma. J Glaucoma. 2022;31:609–613. doi: 10.1097/IJG.0000000000001986. [DOI] [PubMed] [Google Scholar]
- 36.Ekici E., Moghimi S., Hou H., et al. Central visual field defects in patients with distinct glaucomatous optic disc phenotypes. Am J Ophthalmol. 2021;223:229–240. doi: 10.1016/j.ajo.2020.10.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.De Moraes C.G., Prata T.S., Tello C., et al. Glaucoma with early visual field loss affecting both hemifields and the risk of disease progression. Arch Ophthalmol. 2009;127:1129–1134. doi: 10.1001/archophthalmol.2009.165. [DOI] [PubMed] [Google Scholar]
- 38.Kiyota N., Shiga Y., Yasuda M., et al. Sectoral differences in the association of optic nerve head blood flow and glaucomatous visual field defect severity and progression. Invest Ophthalmol Vis Sci. 2019;60:2650–2658. doi: 10.1167/iovs.19-27230. [DOI] [PubMed] [Google Scholar]
- 39.Yousefi S., Sakai H., Murata H., et al. Asymmetric patterns of visual field defect in primary open-angle and primary angle-closure glaucoma. Invest Ophthalmol Vis Sci. 2018;59:1279–1287. doi: 10.1167/iovs.17-22980. [DOI] [PubMed] [Google Scholar]
- 40.Traynis I., De Moraes C.G., Raza A.S., et al. Prevalence and nature of early glaucomatous defects in the central 10 degrees of the visual field. JAMA Ophthalmol. 2014;132:291–297. doi: 10.1001/jamaophthalmol.2013.7656. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Not all clinical data utilized for training or evaluating DeepISP are publicly accessible. For noncommercial purposes, requests to access clinical data about visual tests and related clinical information should be directed to T.N. or E.N. The development codes for DeepISP and synthesizing ISP are available on our GitHub repository (https://github.com/keimy1007/DeepISP).