High-Accuracy Digitization of Humphrey Visual Field Reports Using Convolutional Neural Networks

Abstract

Purpose: Glaucoma is a leading cause of irreversible blindness worldwide, necessitating precise visual field (VF) assessments for effective diagnosis and management. The ability to accurately digitize VF reports is critical for maximizing the utility of the data gathered from clinical evaluations.

Methods: In response to the challenges associated with data accessibility in digitizing VF reports, we developed a lightweight convolutional neural network (CNN) framework. Using a decade-long dataset comprising 15,000 reports, we preprocessed portable document format files and standardized the extracted textual data into 48 × 48 pixel images. To enhance the model's generalization capabilities, we incorporated a variety of font types into the dataset.

Results: The proposed CNN model achieved 100% accuracy in extracting numerical values and over 98.6% accuracy in metadata recognition. Post-processing correction using keyword mapping further improved metadata reliability, effectively addressing errors caused by visually similar characters. The model demonstrated superior efficiency compared to manual data entry, significantly reducing processing time while maintaining near-perfect accuracy.

Conclusions: The findings highlight the effectiveness of our AI-driven digitization method in accurately interpreting Humphrey VF images. This advanced framework provides a reliable solution to digitizing complex visual field reports, thereby facilitating enhanced clinical workflows.

Translational Relevance: The implications of this study extend to streamlined clinical workflows and AI-based report interpretation. By enabling comprehensive trend analysis of visual field changes, our model represents a significant advancement in glaucoma care, showcasing the transformative potential of AI-driven technologies in enhancing precision medicine and improving patient outcomes.

Introduction

Glaucoma stands as a formidable global health challenge, accounting for a substantial portion of irreversible blindness cases worldwide.¹ This complex spectrum of ocular diseases is characterized by the progressive loss of retinal ganglion cells, coupled with structural alterations in the retinal nerve fiber layer and optic nerve head, resulting in a specific pattern of visual field (VF) deficits.² Accurate and timely VF assessments are paramount for the diagnosis, management, and longitudinal monitoring of glaucoma progression. 

The ZEISS Humphrey Field Analyzer (HFA; Carl Zeiss Meditec, Dublin, CA, USA) has long served as the cornerstone of VF testing, earning widespread recognition as the gold standard in clinical practice and research.³ Notably, pivotal glaucoma trials, including the Advanced Glaucoma Intervention Study, the Collaborative Initial Glaucoma Therapy Study, the Early Glaucoma Trial, the Normal Tension Glaucoma Study, and the High Ocular Hypertension Therapy Study, have heavily relied on the Humphrey Perimeter for robust data acquisition and analysis.^4–8 

However, despite its pivotal role, challenges persist regarding data accessibility and compatibility. The proprietary nature of the HFA machine's original data, coupled with the lack of public specifications regarding its database structure, presents significant hurdles. Compounding these challenges is the prevalent storage of perimetry reports in PDF or image files within healthcare organization. Therefore the valuable information garnered from VF assessments remains largely untapped in clinical settings, which otherwise necessitates laborious manual entry of data for research purposes. 

Despite advancements in automating the digitization of Humphrey PDF reports using artificial intelligence (AI), current methodologies predominantly rely on complex object detection models.⁹ Although these models achieve high accuracy, their performance still falls short of near-perfect levels, and their stability remains inadequate for reliably processing all textual elements within the reports, particularly critical data such as raw threshold sensitivity, pattern deviation, and total deviation values. Furthermore, these approaches demand substantial computational power for both training and deployment, posing challenges for practical clinical implementation. In response to these limitations, we conducted this research to develop a lightweight CNN model capable of accurately recognizing all text within Humphrey PDF reports. Our objective is to achieve near-perfect accuracy while ensuring the model's efficiency and feasibility for clinical deployment. 

Material and Methods

Study Design

This is a single-center retrospective study conducted at the Linkou branch of Chang Gung Memorial Hospital, Taiwan. The study received approval from the Chang Gung Medical Foundation Institutional Review Board (IRB no. 202001065B0) and was conducted in accordance with the principles set forth in the Declaration of Helsinki. The study aimed to develop a robust deep learning-based method to digitize Humphrey VF reports, the most widely used perimetric test in clinical practice. A total of 15,000 Humphrey single field analysis (SFA) reports were randomly collected from 2011 to 2021, covering both HFA2 and HFA3 models. These reports, available in PDF format, were used to construct training, validation, and testing datasets for model development. The dataset included reports with different test patterns (30-2, 24-2, and 10-2) and testing strategies (SITA-Standard, SITA-Fast, and SITA-Faster), as detailed in Table 1. 

Table 1.

 

Distribution of Test Patterns and Testing Strategies Across 15,000 Humphrey VF Reports

 

Data Collection and Preprocessing

Humphrey VF reports, available in 30-2, 24-2, and 10-2 SFA PDF formats, underwent meticulous preprocessing to enable accurate data extraction. Detailed analysis revealed that these reports contain 72 unique alphanumeric characters, including numbers, space, symbols, and letters (0123456789 %/+−.,:<>abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ). 

A key observation was that, within PDFs of the same HFA test pattern, the placement of textual and numerical data remains fixed across different reports. This consistency allowed for the development of a coordinate-based extraction framework, where predefined bounding boxes were established based on different HFA formats (HFA2 and HFA3) and test patterns (30-2, 24-2, and 10-2) to efficiently locate and extract relevant text data. This method significantly enhances processing efficiency compared to object detection models, which require additional computational resources for dynamic text localization. Figure 1 illustrates an example of HFA3 24-2 test pattern, where bounding boxes precisely enclose metadata and numerical values, forming a structured coordinate array for text extraction. 

Figure 1.

 

An example of HFA3 24-2 test pattern image, where predefined bounding boxes are used to enclose fixed-position metadata and numerical data, creating a structured coordinate array for data extraction.

Image Cropping and Processing

A total of 15,000 Humphrey VF reports (PDF format) were utilized for model training, validation, and testing, covering 30-2, 24-2, and 10-2 test patterns. Based on the predefined bounding boxes, each individual alphanumeric character was extracted and standardized to 48 × 48 pixels for model input (Fig. 2). 

Figure 2.

 

Schematic diagram illustrates the coordinate array methodology for precise text localization within Humphrey visual field PDFs.

However, challenges were encountered in cases where symbols overlapped with numerical values, particularly triangular markers adjacent to threshold values, which complicated character recognition (Fig. 3). To address this issue, we integrated the iText for .NET open-source library,¹⁰ a robust tool for PDF parsing and modification, to systematically remove non-textual elements such as graphical markers and axis labels, thereby ensuring clean and reliable text extraction. Finally, each cropped character was converted into a 48 × 48-pixel image and incorporated into the dataset. After this refinement, the dataset was manually labeled into 72 distinct character classes, a labor-intensive process that constituted the most time-consuming step of data preprocessing. 

Figure 3.

 

Demonstration of text data overlaps with extraneous markers in Humphrey visual field PDFs, highlighting the necessity of marker and axis removal for improved recognition accuracy using the iText library.

Augmentation for Model Generalization

To enhance model generalization and accommodate varying text fonts, diverse font types were incorporated into the dataset. Alongside images of text extracted from Humphrey PDFs, 50 built-in fonts from Microsoft Windows were used to generate additional training samples. Each font image was standardized to 48 × 48 pixels, resulting in a comprehensive dataset comprising 11,736,000 text images across 72 categories (Fig. 4). 

Figure 4.

 

Representation of the comprehensive dataset, comprising images extracted from Humphrey PDFs and generated font images, facilitating model generalization across diverse text fonts.

Figure 5 provides a comprehensive overview of the data collection and preprocessing workflow, outlining the entire pipeline—from the removal of specific symbols and conversion of PDF files to JPG format, to character cropping, labeling, and data augmentation—ultimately generating the final dataset for model training and evaluation. 

Figure 5.

 

A flowchart illustrating the data processing pipeline for generating the training, validation, and testing datasets.

Model Architecture and Training

A lightweight four-layer CNN architecture was designed for text recognition training (Fig. 6).¹¹ The CNN model, tailored for 48 × 48-pixel input images, comprised successive convolutional layers with 3 × 3 kernels, culminating in 72 output categories corresponding to the recognized characters. The dataset was divided into an 80% training set and a 20% testing set. During each epoch of training, 20% of the training set was randomly sampled and used as the validation set. Model performance was evaluated using fivefold cross-validation to ascertain accuracy and robustness. 

Figure 6.

 

Illustration of the four-layer CNN architecture used for text recognition training, showcasing the sequential convolutional layers and final output categories.

Results

Model Training and Performance Evaluation

After 50 epochs of training on the designated training set, the model demonstrated robust convergence, as evidenced by the convergence of loss values (Fig. 7). During the training phase, accuracy on the validation set reached 99.61%, indicative of the model's proficiency in generalizing to unseen data. On evaluation on the independent testing set, the model achieved an accuracy of 99.67%. A confusion matrix provided a comprehensive assessment of the model's predictive capabilities across 72 character categories, revealing a concentrated distribution along the diagonal, reflecting high accuracy in character recognition tasks (Fig. 8). 

Figure 7.

 

Visualization of accuracy and loss trends over training epochs for both training and validation datasets, demonstrating the model's convergence and performance stability.

Figure 8.

 

Confusion matrix depicting the model's predictive performance across 72 character categories, highlighting the high accuracy and minimal misclassification rates.

Model Performance Evaluation Across Different HFA Formats

To evaluate the model's accuracy across different HFA formats, we assessed performance on HFA2 and HFA3 models, using the 30-2, 24-2, and 10-2 test patterns. The extracted data was categorized into metadata and numerical values, with separate accuracy calculations for each, as summarized in Table 2. 

Table 2.

 

Accuracy of Text Recognition Across Different HFA Models and Test Patterns, Categorized Into Metadata and Numerical Data

 

Our results show that the model consistently achieved 100% accuracy in extracting critical numerical data, including raw threshold sensitivity, total deviation, and pattern deviation values, regardless of HFA model or test pattern. For metadata extraction, the model demonstrated exceptionally high accuracy, exceeding 98.6% across all test formats, with only minimal errors. 

Post Processing

The proposed model is designed for single-character recognition. However, certain characters are inherently difficult to distinguish, even for human eyes, due to their visual similarity. Notable examples include “O” (Orange), “o” (owl), and “0” (zero), as well as “I” (Ice) and “l” (lamp). These similarities frequently lead to misclassification, particularly in metadata fields. 

Although metadata recognition accuracy exceeded 98.6%, a small number of errors persisted, primarily because of the misclassification of these visually similar characters. To address this issue, we implemented a post-processing correction mechanism based on keyword mapping. A predefined dictionary was constructed by identifying commonly occurring metadata keywords and enumerating all potential misrecognized variations. This dictionary was then applied to systematically correct misclassified metadata entries, ensuring accurate keyword restoration. 

Table 3 summarizes the implemented keyword correction strategy. This post-processing approach effectively resolves character misclassification caused by font-based similarities, thereby ensuring highly reliable metadata extraction. 

Table 3.

 

Correction of Commonly Misrecognized Words

 

Computational Efficiency and Comparison With Manual Annotation

To assess computational efficiency, we randomly selected 1,000 PDFs from the testing dataset and processed them using our trained model for inference, followed by post-processing correction. The experiment was conducted on a desktop computer (Intel i7 CPU, 16GB RAM) running TensorFlow 2.10, using only CPU processing without GPU acceleration. The total processing time was 148 minutes, averaging 8.89 seconds per PDF. 

To verify accuracy, all extracted text was manually reviewed, confirming 100% correctness in text recognition. In contrast, manual annotation of PDFs, while theoretically capable of achieving 100% accuracy, is highly labor-intensive and error prone. Manual transcription introduces a risk of typographical errors and misinterpretation, requiring approximately 20 to 30 minutes per report. Extrapolating from this, transcribing 1000 PDFs manually would require an estimated two months of full-time work, underscoring the substantial efficiency gains achieved through automation. 

Examples of AI-Based Character Recognition in Humphrey SFA Reports

Figures 9 to 12 illustrate the AI-based character recognition process applied to Humphrey SFA reports. Figures 9 and 11 present standard HFA2 and HFA3 PDF documents, respectively, serving as input for the recognition system. Figures 10 and 12 display the corresponding recognized results, where the correctly identified metadata and numerical data are extracted and structured. 

Figure 9.

 

Example of an HFA2 PDF document, serving as an input source for the AI-based recognition model.

Figure 10.

 

Recognized output of an HFA2 PDF document (Fig. 9) after model inference and post-processing, illustrating the successful extraction of metadata and numerical data.

Figure 11.

 

Example of an HFA3 PDF document, serving as an input source for the AI-based recognition model.

Figure 12.

 

Recognized output of an HFA3 PDF document (Fig. 11) after model inference and post-processing, illustrating the successful extraction of metadata and numerical data.

Discussion

This study presents a novel AI-driven approach for digitizing Humphrey SFA reports, providing a highly accurate, efficient, and scalable solution for clinical and research applications. Using a lightweight CNN model, the proposed method achieves 100% accuracy in numerical data extraction and 98.6% accuracy in metadata recognition. Automating this process enables structured data storage, retrieval, and integration, facilitating advancements in automated visual field analysis. 

Existing optical character recognition (OCR) solutions fall into two categories: third-party OCR packages and object detection-based models. Third-party OCR tools, such as TesserOCR, are designed for general text recognition and require no custom training.¹² However, their lack of domain-specific adaptation results in suboptimal accuracy when applied to fixed-layout medical reports. Object detection-based models, although more flexible, require substantial computational resources, extensive fine-tuning, and high-performance hardware.⁹ Their multistep segmentation process also introduces error propagation and increases processing time. 

Our model overcomes these limitations by leveraging predefined text coordinates within fixed-layout Humphrey SFA reports, ensuring precise and efficient recognition. Unlike object detection methods, it does not require separate text localization, reducing computational complexity and improving accuracy. Compared to third-party OCR tools, it is specifically trained on HFA2 and HFA3 reports, incorporating domain-specific data to enhance performance. These optimizations result in superior accuracy and efficiency, making the model well suited for large scale clinical implementation. 

Another advantage of our model is its robust generalization across different font styles. By incorporating 50 built-in Windows fonts alongside Humphrey PDF reports, the model was trained to recognize a broad spectrum of character variations, reducing dependency on specific font configurations. This data augmentation strategy enhances adaptability to font variations across different SFA report versions. 

To further improve metadata extraction, we implemented a post-processing correction mechanism to address misclassifications of visually similar characters. A keyword mapping strategy systematically corrected errors, enhancing metadata reliability and ensuring adaptability to real-world clinical applications where precision is critical. Future directions may include integrating recurrent neural networks with CNNs and using Connectionist Temporal Classification¹³ for text sequence prediction, transitioning the system into a fully end-to-end deep learning model without reliance on post-processing. 

Our deep learning model demonstrates exceptional efficacy in digitizing large volumes of Humphrey PDF reports with near-perfect accuracy. The resulting digitized data offers compact storage solutions and facilitate seamless integration into diverse clinical and research workflows. Moreover, these digitized datasets serve as invaluable resources for advanced applications, including AI-based report interpretation. 

Acknowledgments

Disclosure: S.-S. Shie, None; W.-W. Su, None