Abstract
Periodontitis generates substantial clinical data, yet periodontal and PI charts remain predominantly paper-based, which limits reuse for clinical analysis and decision support. We present ChartOCR, a system that digitizes unified periodontal-plaque forms using template-based alignment, coordinate-based region-of-interest extraction, a fine-tuned TrOCR model, and SHA-256. ChartOCR supports common handwriting conventions and provides an interactive interface with real-time result streaming, in-place correction, and CSV export. We evaluated the system on 30 scanned handwritten charts. Periodontal charts achieved a mean accuracy and mAP of 0.987, while PI charts reached 0.942 accuracy and 0.884 mAP. Most digits performed well, although PPV was lower for higher plaque scores. On a MacBook Pro (M3 Pro), the average processing time per chart was approximately 70 seconds. These results demonstrate that our approach enables accurate, efficient digitization of handwritten dental charts and provides a scalable foundation for periodontal informatics research.
Introduction
Periodontitis is a chronic inflammatory disease that affects the supporting structures of the teeth and remains a growing global burden1. With rapid population aging and increased retention of natural teeth, the prevalence of moderate and severe periodontitis is increasing2, followed by a growing number of patients in need of long-term periodontal care3. As periodontitis rarely resolves after a single intervention4, clinicians must provide Supportive Periodontal Therapy (SPT)5 and continuously monitor clinical parameters to detect early progression6. As a result, this generates large amounts of longitudinal clinical data essential for evaluating treatment outcomes and guiding secondary prevention.
Accurate periodontal charting, including Probing Depth (PD), Gingival Recession (GR), and Bleeding on Probing (BOP) at six sites per tooth, forms the basis of periodontal assessment7. However, collecting and recording the high-dimensional measurements, at up to 168 sites per patient, remains labor-intensive and error-prone. Moreover, the 2017 World Workshop classification system requires clinicians to derive interdental Clinical Attachment Loss (CAL) from PD and GR, integrate radiographic bone loss and tooth loss due to periodontitis, and assign severity (Stage) and risk of progression (Grade)8, further increasing cognitive load on clinicians and risks of documentation errors9. In addition, PD, BOP, and plaque index (PI) require tactile measurements and are typically recorded via manual entry or handwritten paper charts10, 11. While prior studies to automate this process using voice recognition12 and radiographic deep learning13 have shown promises, they remain limited by background noise, the inability to capture soft-tissue parameters14, and the lack of tools that directly digitize handwritten periodontal and PI charts from simple images without specialized hardware15.
To bridge the gap in clinical data digitization, we propose an integrated system leveraging OCR to automatically transform handwritten periodontal and PI charts into structured electronic data. We evaluated the system’s performance by assessing character recognition accuracy and end-to-end processing time. We aim to establish a scalable foundation for high-quality periodontal informatics, further facilitating the secondary use of clinical data for precision dentistry.
Related Work
Researchers proposed a variety of tools for automating clinical entry by digitizing data. By reviewing the existing literature, we explore future research opportunities to gain an understanding of OCR tools for digitizing dental data.
Existing Tools for Digitizing Dental Data
Many existing systems have helped clinicians digitize dental data using mouse-and-keyboard input, touchscreens, or voice interfaces. Early work introduced electronic dental record platforms that enable dental students to document oral health findings via graphical interfaces16, 17. Subsequent studies extended this paradigm to mobile applications that support oral health surveys and clinical data collection during fieldwork18, 19, as well as voice-driven12, 20, and hybrid interfaces that further ease data entry and reduce documentation effort21, 22.
On the other hand, some researchers pursued automated digitization using specialized hardware. For example, digital pens and optical readers were used to convert handwritten post-mortem charts, tooth eruption summaries, and plaque indices into structured data23, 24. Although these systems require less time for data entry, they need additional devices. To address this, Jin et al. applied vision–language models to extract information from Chinese dental case screenshots, pointing toward clinical digitization without additional equipment25. In spite of the advances in OCR for handwritten medical data, no existing system directly digitizes dental examination results from images. Moreover, prior tools handle either periodontal status or plaque index, but none supports simultaneous digitization of both.
Existing OCR Tools for Digitizing Handwritten Medical Documents
Researchers have used OCR to digitize handwritten clinical information in diverse medical contexts. Prior studies presented tools that process medical documents using OCR, such as electronic health record (EHR)-scanned documents26, handwritten doctors’ notes27, and medical documents written in various languages28. However, using OCR alone remains challenging due to the complex layouts and contextual cues embedded in medical documents.
To improve prescription digitization, several studies combined OCR with advanced models or supporting technologies. Examples include integrating CRNNs for prescription and EHR processing29, using Llama-based systems for keyword extraction and risk assessment30, incorporating fog-enabled deep learning for prescription and pathology detection31, and securing digitized prescriptions via blockchain architectures32. Despite the innovations, OCR-only approaches still struggle to capture contextual information inherent in clinical records.
To leverage such context, researchers turned to bounded forms. Previous studies digitized structured ophthalmology forms33, captured handwritten dental entries from EHR interfaces25, and developed OCR-tailored templates for specific clinical documentation tasks24, 34. These systems demonstrate the value of constraining input layouts to improve accuracy. However, none of the prior studies address how to make form completion itself easier for clinicians. In particular, no existing OCR-based handwritten recording system explicitly supports practical conventions such as leaving fields blank to represent zero or using dashes (–) to indicate repeated entries.
Methods
To prepare the data and model for implementing ChartOCR, we went through the following stages: template design, module development, model fine-tuning, and evaluation of the fine-tuned model.
Setup
Periodontal Chart Template Design Figure 1 presents the chart template we developed for our system. This form allows clinicians to complete periodontal and PI charts on a single sheet, with three components: a header section (patient name, examination date, chart ID), a periodontal chart section, and a PI chart section. The periodontal and PI charts are aligned in a column-based layout. Each tooth column has a small box adjacent to the tooth number where clinicians can mark an “x” to indicate a missing tooth. When a tooth is missing, its data are encoded as “88”.
Figure 1.
Periodontal-plaque index chart image used in our system. Both charts support marking missing teeth by leaving an “X” in a designated box, which is illustrated in the example on the bottom right.
For each tooth, clinicians used a standardized UNC probe (Hu-Friedy, Chicago, IL, USA) to measure four parameters, such as PD, GR, BOP, and PI across six sites including mesiobuccal, mid-buccal, distobuccal, mesiolingual, mid-lingual, and distolingual. PD was measured from the gingival margin to the base of the pocket (1–15 mm), and GR from the cementoenamel junction to the gingival margin (0–15 mm). BOP was marked 0 or 1 to indicate the absence or presence of bleeding, and PI was scored from 0 to 5 using the Turesky modification of the Quigley–Hein index35. When entering values, clinicians could leave a box empty to denote 0, and use dashes (–) to substitute number 1s.
Chart Structure Localization With the raw periodontal chart template we designed, we created a fixed-sized chart coordinate template. This template defines the relative positions of key table elements and serves as the geometric basis for later alignment with real chart inputs. To construct it, we converted the reference image to grayscale, applied thresholding to enhance contrast and reveal table lines, and then used morphological and bitwise operations to isolate horizontal and vertical strokes and reconstruct the full grid. With a coherent grid in place, we applied contour detection to identify connected regions, selected the largest contours as the main table areas, and extracted their bounding boxes. Using the same procedure, we then detected individual cell and checkbox regions, computed cell coordinates while removing overlaps via Non-Maximum Suppression (NMS), and analyzed the header or footer with Canny edge detection and contour filtering to localize checkboxes. Finally, we exported all extracted coordinates as a JSON file, where each bounding box is represented as [x, y, width, height] in pixel units.
Module Development
Image Preprocessing Module Once the system receives an input chart image, it deskews the chart using an adaptive radial projection on the Fourier magnitude spectrum36. Following this, it aligns the calibrated image to the template (see Figure 1) by detecting OCR features, matching them, and estimating a homography to warp the input into the template’s coordinate space, thereby ensuring a consistent layout for later steps. The module then exports the aligned image as an intermediate output.
Chart Processing Module The system identifies the periodontal and PI chart regions from the aligned image. These two regions of interest (ROI) correspond to the areas on the chart containing clinical measurements. Each ROI is then divided into individual cells, which contains a single digit. For each cropped cell image, the system attempts to retrieve the corresponding digit value from the cache using SHA-256, assigning it to the cell if a match is found. Finally, the module outputs the digits corresponding to each cell of the chart.
Result Standardization Module We developed a standardization module to convert the module outputs into a format that aligns with real-world clinical records. The module standardizes the raw inputs into a structured format. The standardization process consists of two stages. First, it validates the checkbox values. If a checkbox is marked as “X”, all corresponding cell entries are set to “88” to denote a missing tooth. Otherwise, it normalizes all outputs to a three-character format, replacing overly long entries with “000” and padding shorter ones to the length of 3. Second, it formats the output into structured entries, where each entry contains tooth number, surface, evaluation criteria, and mesial/central/distal measurements.
Model Fine-tuning
Model To identify the optimal model for fine-tuning, we established two primary selection criteria. First, we needed a model pre-trained on multiple handwriting styles because our system aimed to enable multiple clinicians to use handwriting shortcuts when filling out the chart. Second, we needed a model with high handwriting recognition accuracy to prevent information distortion. As a result, we found TrOCR, an OCR model based on the transformer model37. TrOCR was pre-trained on large-scale datasets containing both printed and handwritten text, enabling it to handle diverse handwriting styles1. Also, TrOCR demonstrated high handwriting recognition accuracy compared to other OCR models37. Since TrOCR’s properties meet the two eligibility criteria, TrOCR aligns perfectly with our requirements. Thus, the model was integrated into our system.
Dataset To generate training data, we constructed 50 synthetic periodontal and PI charts whose PD, GR, and BOP values satisfied predefined validity constraints, and sampled 225 tooth entries from them. As shown in Figure 2, each tooth entry contains six numeric fields. In this pilot implementation, we constrained each site measurement from 0 to 9. From these entries, we manually cropped two three-field cell images for each entry, which resulted in a total of 550 cell images. Once we assigned images to their ground-truth labels, we resized each input image to 384 x
Figure 2.
Workflow for preparing a fine-tuning dataset from paper-based periodontal charts. From the input chart (1), we manually selected tooth entries (2), cropped each three-character cell image (3), and assigned ground-truth labels based on the handwritten values (4).
384. Additionally, to improve the model’s robustness across varied lighting conditions and image quality, we applied multiple augmentations to the images, including random zooming, rotation, and exposure adjustment. This process expanded the final dataset to 880 images, which we then split into training, validation, and test subsets at a ratio of 75:13:13.
Fine-tuning We trained the model using the datasets constructed in the previous stage. To determine the appropriate learning rate for training, we evaluated three learning rates: 1e−5, 2e−5, and 5e−5 (batch size 8; 6 epochs; validation every 50 steps). As shown in Table 1, the learning rate of 5e−5 achieved the lowest CER. Therefore, we performed fine-tuning using 5e−5 for our OCR model.
Table 1.
CER Result with three learning rates – 1e−5, 2e−5, and 5e−5
| Learning rate | 1e−5 | 2e−5 | 5e−5 |
|---|---|---|---|
| CER on test set | 0.0785 | 0.1126 | 0.0340 |
Model Evaluation
Dataset To evaluate the performance of the fine-tuned model, we made and scanned 30 synthetic handwritten charts, with 10 per periodontitis stage, handwritten by two dental hygienists, SR and HH (15 charts each). The periodontitis stage is determined by CAL, computed as the sum of PD and GR at each site. Following established criteria8, we categorized each chart as Stage I (1–2 mm), Stage II (3–4 mm), or Stages III–IV (≥ 5 mm) according to the highest CAL in the value. The statistics of these charts are summarized in Table 2.
Table 2.
Summary of periodontal clinical data (n = 30). For each image, the table reports bleeding on probing (BOP,%), probing depth (PD, mm), gingival recession (GR, mm), number of missing teeth, clinical attachment level (CAL, mm), and disease stage. PD, GR, and CAL are presented as mean ± standard deviation.
| Image | BOP (%) | PD (mm) | GR (mm) | CAL (Av.) | CAL (MAX) | Missing Teeth (n) | Stage |
|---|---|---|---|---|---|---|---|
| P1 | 0.6 | 1.2 ± 0.4 | 0.0 ± 0.1 | 1.2 ± 0.4 | 2 | 0 | I |
| P2 | 1.2 | 1.1 ± 0.3 | 0.0 ± 0.1 | 1.1 ± 0.3 | 2 | 0 | I |
| P3 | 0.0 | 1.1 ± 0.4 | 0.0 ± 0.0 | 1.1 ± 0.4 | 2 | 2 | I |
| P4 | 0.0 | 1.1 ± 0.3 | 0.0 ± 0.0 | 1.1 ± 0.3 | 2 | 0 | I |
| P5 | 0.0 | 1.1 ± 0.4 | 0.0 ± 0.0 | 1.1 ± 0.4 | 2 | 1 | I |
| P6 | 0.0 | 1.1 ± 0.3 | 0.0 ± 0.0 | 1.1 ± 0.3 | 2 | 0 | I |
| P7 | 0.0 | 1.2 ± 0.4 | 0.0 ± 0.0 | 1.2 ± 0.4 | 2 | 0 | I |
| P8 | 2.3 | 1.6 ± 0.5 | 0.0 ± 0.0 | 1.6 ± 0.5 | 2 | 0 | I |
| P9 | 2.3 | 1.4 ± 0.5 | 0.0 ± 0.0 | 1.4 ± 0.5 | 2 | 0 | I |
| P10 | 5.4 | 1.8 ± 0.4 | 0.0 ± 0.0 | 1.8 ± 0.4 | 2 | 0 | I |
| P11 | 1.8 | 1.3 ± 0.5 | 0.0 ± 0.2 | 1.4 ± 0.5 | 3 | 0 | II |
| P12 | 1.2 | 1.2 ± 0.6 | 0.0 ± 0.1 | 1.2 ± 0.7 | 3 | 3 | II |
| P13 | 0.0 | 1.3 ± 0.5 | 0.0 ± 0.1 | 1.3 ± 0.5 | 3 | 0 | II |
| P14 | 3.0 | 1.1 ± 0.4 | 0.0 ± 0.1 | 1.1 ± 0.4 | 4 | 0 | II |
| P15 | 4.8 | 1.5 ± 0.5 | 0.1 ± 0.5 | 1.6 ± 0.7 | 4 | 0 | II |
| P16 | 6.0 | 1.9 ± 0.9 | 0.1 ± 0.4 | 2.0 ± 1.0 | 4 | 2 | II |
| P17 | 10.1 | 1.7 ± 0.6 | 0.0 ± 0.2 | 1.8 ± 0.6 | 3 | 0 | II |
| P18 | 23.8 | 1.3 ± 0.7 | 0.0 ± 0.1 | 1.4 ± 0.7 | 3 | 1 | II |
| P19 | 4.8 | 1.1 ± 0.5 | 0.1 ± 0.2 | 1.2 ± 0.6 | 3 | 1 | II |
| P20 | 7.1 | 1.2 ± 0.5 | 0.0 ± 0.2 | 1.2 ± 0.5 | 3 | 1 | II |
| P21 | 6.0 | 1.6 ± 0.8 | 0.1 ± 0.4 | 1.7 ± 0.9 | 5 | 0 | III |
| P22 | 8.3 | 2.3 ± 0.8 | 0.1 ± 0.4 | 2.4 ± 0.9 | 5 | 0 | III |
| P23 | 8.3 | 1.4 ± 0.7 | 0.7 ± 0.1 | 2.1 ± 1.4 | 6 | 2 | III |
| P24 | 11.9 | 1.7 ± 0.9 | 0.2 ± 0.6 | 1.9 ± 1.2 | 5 | 2 | III |
| P25 | 13.7 | 1.6 ± 0.8 | 0.5 ± 0.8 | 2.1 ± 1.2 | 5 | 0 | III |
| P26 | 27.4 | 1.3 ± 0.6 | 0.0 ± 0.4 | 1.4 ± 0.7 | 5 | 1 | III |
| P27 | 10.7 | 1.3 ± 0.6 | 0.9 ± 0.8 | 1.7 ± 1.0 | 6 | 1 | III |
| P28 | 22.0 | 2.7 ± 1.5 | 0.6 ± 0.8 | 3.3 ± 1.7 | 8 | 0 | III |
| P29 | 15.5 | 2.9 ± 1.6 | 1.1 ± 0.8 | 4.0 ± 1.8 | 8 | 0 | III |
| P30 | 24.4 | 3.4 ± 1.6 | 0.4 ± 0.7 | 3.8 ± 1.7 | 8 | 0 | III |
Metrics We processed the 30 test images constructed in the previous stage using our model and evaluated its performance using three metrics: (1) Accuracy, defined as the ratio of correctly recognized cells to the total number of cells; (2) mean Average Precision (mAP), which reflects the model’s robustness in recognizing digits under varied chart layouts; and (3) Processing Time, the average time required for our model to process one chart image. All experiments were conducted on a MacBook Pro equipped with an M3 Pro chip.
Furthermore, we examined whether specific digits tended to underperform across periodontal and PI charts. For each digit appearing in the chart, we computed its True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). From these counts, we derived four per-digit evaluation metrics: (1) Positive Predicted Value (PPV), indicating how often predicted positives are truly positive; (2) Negative Predicted Value (NPV), representing the reliability of predicted negatives; (3) Sensitivity (Sn), measuring how effectively the model retrieves actual positives; and (4) Specificity (Sp), indicating how well it distinguishes true negatives.
System Implementation
We implemented ChartOCR, an end-to-end OCR system for digitizing periodontal and PI charts. This system sequentially performs image deskewing and alignment, coordinate-guided ROI extraction, SHA-256, and batched inference through a fine-tuned TrOCR inference engine (see Figure 3). To support real-world clinical workflows, the system streams intermediate results to the frontend, provides in-place correction with real-time validation, and exports standardized outputs for downstream analysis.
Figure 3.
Overview of the ChartOCR. The system (1) deskews the scanned chart and aligns it with the template image to obtain a normalized layout. (2) For each ROI, it retrieves cached results when available or performs OCR inference. (3) Raw outputs are standardized per tooth entry by checking missing-tooth markings and formatting recognized digits into the required 3-character structure.
System Initialization When the system boots, it first loads the coordinate template file and initializes the ChartOCR engine on the available hardware: GPU (for CUDA-enabled devices), MPS (for Mac devices with Apple Silicon chips), or CPU. Built on top of our fine-tuned TrOCR model, this engine supports batched inference and SHA-256 mechanism to minimize redundant computation and latency during OCR processing.
Backend Design We hosted our backend using FastAPI2 for high performance and asynchronous request handling. We implemented two POST endpoints to support the application’s functionality. The first endpoint, /ocr, receives uploaded images, runs the OCR pipeline, and streams intermediate results so users can track progress. The second endpoint, /export, packages periodontal and plaque outputs as CSV files in a ZIP archive for downstream data management and analysis.
Frontend Design We designed our user interface with React3 and Tailwind CSS4 to create an interface that reduces cognitive load for clinicians, provides immediate visual feedback, and keeps the entire digitization workflow transparent while the OCR engine operates in the background. To accomplish this, we organized the workflow into two primary screens: the Home screen and the Result screen. The Home screen offers drag-and-drop upload and manual file selection via a “Click to Upload” button for JPG, PNG, and BMP files up to 10 MB. After a upload, the system confirms success via visual feedback, enables the “Start Processing” button, and shows an image preview that users can replace or remove before initiating OCR.
During processing, the application displays a split-screen layout with the input image on the left and a streaming result table on the right. A status indicator reports the current stage and disappears when the pipeline finishes. The system displays periodontal results as soon as they are ready while continuing to process plaque data, and disables Save and Export until all outputs are complete.
In View mode, the results are organized into periodontal and plaque tabs (see Figure 4). Each row displays the tooth number, missing-tooth status (green tag if not missing, red otherwise), surface type (B for buccal, L for labial/lingual, P for palatal), evaluation criteria, and mesial/central/distal measurements. In Edit mode, users are allowed to modify measurement fields and missing-tooth labels, with real-time validation and contextual tooltips for incorrect inputs (e.g., “BOP must be 0 or 1”). After confirming edits, users can save the changes and export the final ZIP file.
Figure 4.
Result screen of ChartOCR. (a) Users can view their uploaded chart image on the left panel. (b) The system streams OCR outputs in an interactive table, allowing users to scroll, inspect, and compare values across tooth surfaces. (c) When Edit mode is enabled, users are allowed to directly correct OCR values before finalizing the results. (d) In View mode, users can return to the home screen or export the finalized results as a ZIP file.
Result
Table 3 presents the overall system performance by chart type, periodontitis stage, and full-image processing time. For periodontal charts, Stage I achieved the highest chart-level accuracy and mAP, followed by Stage II and III–IV. Across all stages, accuracy ranged from 0.921 and 1 (mean 0.987), and mAP from 0.895 to 1 (mean 0.987). PI charts follow a similar pattern, with Stage I showing significantly higher accuracy and mAP than the other two stages. Overall accuracy ranged from 0.845 to 1 (mean 0.942), and mAP from 0.542 to 1 (mean 0.884). The mAP value was relatively lower in the PI chart, and this tendency became more pronounced as the stage number increased. For full-image processing, the per-image processing time ranged from 68.879 to 71.584 seconds, with an average of 70.140 seconds.
Table 3.
Overall system performance (rounded by 3 decimal points)
| Target | Stage | Metric | Min | Max | Average |
|---|---|---|---|---|---|
| Periodontal chart | Stage I | Accuracy (%) | 0.994 | 1.000 | 0.998 |
| mAP (%) | 0.972 | 1.000 | 0.995 | ||
| Stage II | Accuracy (%) | 0.972 | 1.000 | 0.992 | |
| mAP (%) | 0.909 | 1.000 | 0.989 | ||
| Stages III–IV | Accuracy (%) | 0.921 | 0.998 | 0.970 | |
| mAP (%) | 0.895 | 0.999 | 0.976 | ||
| All stages | Accuracy (%) | 0.921 | 1.000 | 0.987 | |
| mAP (%) | 0.895 | 1.000 | 0.987 | ||
| Plaque index chart | Stage I | Accuracy (%) | 0.964 | 1.000 | 0.979 |
| mAP (%) | 0.726 | 1.000 | 0.930 | ||
| Stage II | Accuracy (%) | 0.887 | 0.994 | 0.930 | |
| mAP (%) | 0.663 | 1.000 | 0.859 | ||
| Stages III–IV | Accuracy (%) | 0.845 | 1.000 | 0.916 | |
| mAP (%) | 0.542 | 1.000 | 0.851 | ||
| All stages | Accuracy (%) | 0.845 | 1.000 | 0.942 | |
| mAP (%) | 0.542 | 1.000 | 0.884 | ||
| Full image | Processing time | Time (seconds) | 68.879 | 71.584 | 70.140 |
Table 4 provides a breakdown of per-digit PPV, NPV, Sn, and Sp for periodontal and PI charts. For the periodontal chart, the missing-value annotation X achieved perfect scores for all metrics, and digits 0–3 showed consistently strong performance (PPV, NPV, Sn, Sp ≥ 0.98). Digits 4–6 maintained high NPV and Sp (≥ 0.96), but exhibited lower PPV (0.780–0.989) and Sn (0.316–0.693). For the PI chart, X again achieved perfect PPV and Sp, with NPV of 0.996 and Sn of 0.944. Digits 1–3 performed robustly (PPV 0.903–0.952, NPV 0.972–0.996, Sn 0.859–0.989, Sp 0.973–0.989), whereas digits 0, 4, and especially 5 showed lower PPV (0.785, 0.875, and 0.256, respectively) despite retaining high NPV (0.964–0.999) and Sp (0.960–1). Digits 7–9 in the periodontal chart are not reported because they did not appear in the evaluation set.
Table 4.
Per-digit metric performance (rounded by 3 decimal points)
| (a) Periodontal chart per-digit performance | ||||
|---|---|---|---|---|
| Class | PPV (%) | NPV (%) | Sn (%) | Sp (%) |
| X | 1.000 | 1.000 | 1.000 | 1.000 |
| 0 | 0.991 | 0.998 | 0.998 | 0.991 |
| 1 | 0.988 | 0.996 | 0.986 | 0.997 |
| 2 | 0.986 | 0.999 | 0.989 | 0.998 |
| 3 | 0.990 | 0.995 | 0.919 | 1.000 |
| 4 | 0.989 | 0.992 | 0.693 | 0.999 |
| 5 | 0.780 | 0.993 | 0.593 | 0.997 |
| 6 | 0.981 | 0.965 | 0.316 | 0.999 |
| 7–9 | – | – | – | – |
| (b) Plaque index chart per-digit performance | ||||
|---|---|---|---|---|
| Class | PPV (%) | NPV (%) | Sn (%) | Sp (%) |
| X | 1.000 | 0.996 | 0.944 | 1.000 |
| 0 | 0.785 | 0.964 | 0.672 | 0.979 |
| 1 | 0.922 | 0.986 | 0.975 | 0.988 |
| 2 | 0.952 | 0.996 | 0.989 | 0.973 |
| 3 | 0.903 | 0.972 | 0.859 | 0.989 |
| 4 | 0.875 | 0.964 | 0.700 | 1.000 |
| 5 | 0.256 | 0.999 | 0.750 | 0.960 |
Discussion
We developed and validated ChartOCR digitizing handwritten periodontal and PI charts using OCR. By enabling clinicians to record findings on a unified form and instantly convert them into structured data, the system bridges traditional paper-based workflows with data-driven periodontal management. Evaluation on 30 handwritten charts demonstrated high accuracy across all periodontitis stages, indicating that reliable digitization is achievable without specialized hardware. To our knowledge, this is the first OCR-based system capable of simultaneously digitizing periodontal and PI charts, and one of the first attempts to extract dental examination results directly from handwritten images. Prior OCR-based tools did not address practical charting conventions, such as blank fields for zero or dashes (–) for repeated entries, whereas our system integrates these rules into both form design and the recognition pipeline. This work fills critical gaps in dental informatics, particularly the lack of tools supporting realistic annotation strategies.
Our study has several limitations. First, while we combined the periodontal and PI charts to detect early signs of disease progression, the combined chart does not integrate radiographic indicators, such as furcation involvement and tooth-loss tracking, which may lead to staging underestimation. Second, we did not distributed digits uniformly during fine-tuning process but focused on collecting the low-value digits appearing frequently to implement a functional prototype at a low cost. Third, while the TrOCR backbone was chosen for its ability to handle diverse handwriting styles, data for fine-tuning a model were created by only two handwriting contributors, which potentially limits generalizability. Last, we did not evaluate the usability and feasibility of the system with stakeholders, such as dentists and other researchers, although the fine-tuned model was evaluated using multiple standard metrics (e.g., mAP, PPV, NPV, Sn, and Sp).
Future work remains to embed direct clinician feedback, broaden handwriting diversity, and expand the range of supported periodontal chart types and encoding. It might be valuable to employ grid-based data reduction38, synthetic data augmentation, and expanded clinical datasets to improve the sensitivity of the high-value digits. Furthermore, it is essential to evaluate the model’s performance on real-world multi-center data to establish its robustness and clinical readiness. Such efforts should ultimately culminate in large-scale studies involving diverse patient cohorts to verify the system’s effectiveness and reliability in clinical practice.
Conclusion
Our proposed system streamlines periodontal documentation by automatically structuring 168 clinical data points from scanned chart images, eliminating the inefficiencies of manual entry. This study offers a practical solution for transforming analog clinical records into high-quality digital assets, establishing a foundation for the future implementation of data-driven precision dentistry.
Acknowledgement
This work was supported by the New Faculty Startup Fund from Seoul National University. This work was also supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2024-00407105).
Footnotes
Figures & Tables
References
- 1.Cekici A, Kantarci A, Hasturk H, Van Dyke TE. Inflammatory and immune pathways in the pathogenesis of periodontal disease. Periodontol 2000. 2014 Feb;64(1):57–80. doi: 10.1111/prd.12002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Kassebaum NJ, Bernabé E, Dahiya M, Bhandari B, Murray CJL, Marcenes W. Global burden of severe periodontitis in 1990-2010: a systematic review and meta-regression. J Dent Res. 2014 Sep;93(11):1045–53. doi: 10.1177/0022034514552491. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Tonetti MS, Jepsen S, Jin L, Otomo-Corgel J. Impact of the global burden of periodontal diseases on health, nutrition and wellbeing of mankind: A call for global action. J Clin Periodontol. 2017 May;44(5):456–62. doi: 10.1111/jcpe.12732. [DOI] [PubMed] [Google Scholar]
- 4.Rattu V, Patel T, Loke J, Petsos H, Nibali L. Long-term clinical benefits of periodontal interventions in strict supportive periodontal care: A systematic review. J Periodontol. 2025 Nov. [DOI] [PMC free article] [PubMed]
- 5.Hirschfeld L, Wasserman B. A long-term survey of tooth loss in 600 treated periodontal patients. J Periodontol. 1978 May;49(5):225–37. doi: 10.1902/jop.1978.49.5.225. [DOI] [PubMed] [Google Scholar]
- 6.Lindhe J, Nyman S. Long-term maintenance of patients treated for advanced periodontal disease. J Clin Periodontol. 1984 Sep;11(8):504–14. doi: 10.1111/j.1600-051x.1984.tb00902.x. [DOI] [PubMed] [Google Scholar]
- 7.Eke PI, Page RC, Wei L, Thornton-Evans G, Genco RJ. Update of the case definitions for population-based surveillance of periodontitis. J Periodontol. 2012 Mar;83(12):1449–54. doi: 10.1902/jop.2012.110664. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Tonetti MS, Greenwell H, Kornman KS. Staging and grading of periodontitis: Framework and proposal of a new classification and case definition. J Periodontol. 2018 Jun;89(1):S159–72. doi: 10.1002/JPER.18-0006. [DOI] [PubMed] [Google Scholar]
- 9.Dietrich T, Ower P, Tank M, West NX, Walter C, Needleman I, et al. Periodontal diagnosis in the context of the 2017 classification system of periodontal diseases and conditions - implementation in clinical practice. Br Dent J. 2019 Jan;226(1):16–22. doi: 10.1038/sj.bdj.2019.3. [DOI] [PubMed] [Google Scholar]
- 10.Clark-Perry D, Van der Weijden GA, Berkhout WER, Wang T, Levin L, Slot DE. ACCURACY OF CLINICAL AND RADIOGRAPHIC MEASUREMENTS OF PERIODONTAL INFRABONY DEFECTS OF DIAGNOSTIC TEST ACCURACY (DTA) STUDIES: A SYSTEMATIC REVIEW AND META-ANALYSIS. J Evid Based Dent Pract. 2021 Nov;22(1):101665. doi: 10.1016/j.jebdp.2021.101665. [DOI] [PubMed] [Google Scholar]
- 11.Heitz-Mayfield LJA. Conventional diagnostic criteria for periodontal diseases (plaque-induced gingivitis and periodontitis) Periodontol 2000. 2024 Jun;95(1):10–9. doi: 10.1111/prd.12579. [DOI] [PubMed] [Google Scholar]
- 12.Nagy M, Hanzlicek P, Zvarova J, Dostalova T, Seydlova M, Hippman R, et al. Voice-controlled data entry in dental electronic health record. Stud Health Technol Inform. 2008;136:529–34. [PubMed] [Google Scholar]
- 13.Krois J, Ekert T, Meinhold L, Golla T, Kharbot B, Wittemeier A, et al. Deep Learning for the Radiographic Detection of Periodontal Bone Loss. Sci Rep. 2019 Jun;9(1):8495. doi: 10.1038/s41598-019-44839-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Ferrara E, Rapone B, D’Albenzio A. Applications of deep learning in periodontal disease diagnosis and management: a systematic review and critical appraisal. Journal of Medical Artificial Intelligence. 2024;8(0) [Google Scholar]
- 15.Khanagar SB, Al-Ehaideb A, Maganur PC, Vishwanathaiah S, Patil S, Baeshen HA, et al. Developments, application, and performance of artificial intelligence in dentistry - A systematic review. J Dent Sci. 2020 Jun;16(1):508–22. doi: 10.1016/j.jds.2020.06.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Suebnukarn S, Rittipakorn P, Thongyoi B, Boonpitak K, Wongsapai M, Pakdeesan P. Usability assessment of an electronic health record in a comprehensive dental clinic. Springerplus. 2013 May;2(1):220. doi: 10.1186/2193-1801-2-220. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Lacasandile AD, Tiu GV, Victoria NM, De Lemos AN, Era AD. National University Dental Records Electronic Access Management (NU-DREAM) as an Electronic Dental Record in a University Setting. 2024 6th International Workshop on Artificial Intelligence and Education (WAIE) 2024:p.265–71. [Google Scholar]
- 18.Detsomboonrat P, Pisarnturakit PP. Time Efficiency, Reliability, and User Satisfaction of the Tooth Memo App for Recording Oral Health Information: Cross-Sectional Questionnaire Study. JMIR Form Res. 2024 Apr;8:e56143. doi: 10.2196/56143. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Abdelrahman HH, Hamza M, Essam W, Adham M, AbdulKafi A, Baniode M. Electronic oral health surveillance system for Egyptian preschoolers using District Health Information System (DHIS2): design description and time motion study. BMC Oral Health. 2024 Jul;24(1):807. doi: 10.1186/s12903-024-04550-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Hippmann R, Dostálová T, Zvárová J, Nagy M, Seydlova M, Hanzlícek P, et al. Voice-supported electronic health record for temporomandibular joint disorders. Methods Inf Med. 2009 Nov;49(2):168–72. doi: 10.3414/ME9301. [DOI] [PubMed] [Google Scholar]
- 21.Preissner S, Kostka E, Mokross M, Kersten NV, Blunck U, Preissner R. DBEndo: a web-based endodontic case management tool. BMC Res Notes. 2015 Nov;8:685. doi: 10.1186/s13104-015-1680-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Zvárová J, Chleborád K, Zvára K, Dostálová T. Medical Informatics and Information Technology Supporting Oral Medicine. Stud Health Technol Inform. 2016;223:230–6. [PubMed] [Google Scholar]
- 23.Ohira H, Yamamoto I, Sadamori T, Miyauchi H, Nakagawa K, Hukushige H, et al. Transmission of dental data on a microdotted paper-based dental chart from a digital pen – trial operations at disaster exercise sites and at a landslide disaster at Hiroshima City [Article] Journal of the Canadian Society of Forensic Science. 2016;49(4):189–202. [Google Scholar]
- 24.Helm S. Recording system for the Danish Child Dental Health Services. Community Dent Oral Epidemiol. 1973;1(1):3–8. doi: 10.1111/j.1600-0528.1973.tb01055.x. [DOI] [PubMed] [Google Scholar]
- 25.Jin W, Sun Y, Ji K, Jiang X, Hu Y, Wang J, et al. MedScreenDental: Automated structured dental record generation via multimodal language model integration. Displays. 2025;90:103119. [Google Scholar]
- 26.Kostrinsky-Thomas AL, Hisama FM, Payne TH. Searching the PDF haystack: Automated knowledge discovery in scanned EHR documents. Appl Clin Inform. 2021 Mar;12(2):245–50. doi: 10.1055/s-0041-1726103. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Chandra R, Tiwari S, Agarwal S, Singh N. Semantic web-based diagnosis and treatment of vector-borne diseases using SWRL rules. Know-Based Syst. 2023 Aug;274(C) [Google Scholar]
- 28.Prashanth K, Kowndinya B, Vijay C, Teja D, Rodge V, Velaga R, et al. A Platform for Large Scale Auto Annotation of Scanned Documents Featuring Real-Time Model Building and Model Pooling. In: Raman B, Murala S, Chowdhury A, Dhall A, Goyal P, editors. Computer Vision and Image Processing. Cham: Springer International Publishing; 2022. pp. p.58–70. [Google Scholar]
- 29.R N, Priya S K, P RK, Arun M D, N S, R N, et al. Medical Assistance System Using Large Language Model and Medical Data Management. 2025 3rd International Conference on Advancements in Electrical, Electronics, Communication, Computing and Automation (ICAECA) 2025:p.1–5. [Google Scholar]
- 30.Verma R, Pandey S, Awasthi S, Shukla S. Simplifying Medical Report: A Novel Approach to Medical Reporting Using OCR Technology. 2025 3rd International Conference on Communication, Security, and Artificial Intelligence (ICCSAI) 2025;3:p.2077–82. [Google Scholar]
- 31.Drabo C, Malo S. Fog-Enabled Modular Deep Learning Platform for Textual Data Mining in Healthcare for Pathology Detection in Burkina Faso. Stud Health Technol Inform. 2025 Jun;328:173–7. doi: 10.3233/SHTI250696. [DOI] [PubMed] [Google Scholar]
- 32.Malathi P, Sai BC, Sahithi B, Sravani KV, Vamsi PT. Blockchain-Driven Healthcare Management: A Secure and Efficient Prescription System with OCR and QR Code. In: Patnaik LM, G K R, Prasad NH, editors. Advances in Communication and Applications. Singapore: Springer Nature Singapore; 2025. pp. p.77–88. [Google Scholar]
- 33.Rasmussen LV, Peissig PL, McCarty CA, Starren J. Development of an optical character recognition pipeline for handwritten form fields from an electronic health record. J Am Med Inform Assoc. 2011 Sep;19(e1):e90–5. doi: 10.1136/amiajnl-2011-000182. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Mulcahy JJ. AI-Driven Document Intelligence for Integrating EHR and RCM Systems During Hospital System Outages: A Case Study. 2025 IEEE International systems Conference (SysCon) 2025:p.1–7. [Google Scholar]
- 35.Turesky SS, Gilmore ND, Glickman I. Reduced plaque formation by the chloromethyl analogue of victamine C. Journal of periodontology. 1970;41 1:41–3. doi: 10.1902/jop.1970.41.41.41. [DOI] [PubMed] [Google Scholar]
- 36.Pham L, Hoang PH, Mai XT, Tran TA. Adaptive Radial Projection on Fourier Magnitude Spectrum for Document Image Skew Estimation. 2022 IEEE International Conference on Image Processing (ICIP) 2022:p.1061–5. [Google Scholar]
- 37.Li M, Lv T, Chen J, Cui L, Lu Y, Florencio D, et al. TrOCR: Transformer-Based Optical Character Recognition with Pre-trained Models. Proceedings of the AAAI Conference on Artificial Intelligence. 2023 Jun;37(11):13094–102. Available from: https://ojs.aaai.org/index.php/AAAI/article/view/26538 . [Google Scholar]
- 38.Fabrice N, Gang S, Lee JJ. Training Data Sets Construction from Large Data Set for PCB Character Recognition. Journal of Multimedia Information System. 2019;6(4):225–34. [Google Scholar]