Validation of a You Only Look Once–LabelMe Hybrid AI Model for Automated Assessment of Pressure Injury Size and Pocket

ABSTRACT

Quantitative assessment of the pressure injury size (S) and pocket (P, undermining) is essential for evaluating wound healing. Conventional LabelMe annotations often lead to oversegmentation, whereas object detection with You Only Look Once (YOLO) enables accurate wound localization. Here, we developed a hybrid AI model that combines LabelMe segmentation with YOLO gating. This single‐centre retrospective study compared two models: LabelMe‐Seeded Auto‐Recognition (LSAR) and YOLO‐gated LSAR (YGL). Model performance was evaluated by concordance rate and weighted Cohen's κ for DESIGN‐R staging and by mean absolute error (MAE) and median (interquartile range) for area error. Overall, 1017 wound images (training: 979; testing: 38) were analysed. For S, the YGL model achieved higher concordance (85.7%, κ = 0.969) and smaller area errors than did the LSAR. For P, the YGL model demonstrated higher concordance and smaller area errors than did the LSAR model. The YOLO–LabelMe hybrid model improved the stage concordance and size accuracy compared with LabelMe alone. Despite the residual outliers in P assessment, this approach represents a promising step towards automated, clinically adaptable wound measurements.

Key Points

• A hybrid AI model combining LabelMe polygonal segmentation with YOLO‐based gating (YGL) was developed to automatically calculate the size (S) and pocket (P) components of the DESIGN‐R pressure injury assessment scale from clinical photographs.

• For size, YOLO gating suppressed over‐recognition of non‐wound regions and improved stage concordance from 61.9% to 85.7% (weighted κ, 0.492 to 0.969), while reducing the mean absolute area error from 15.22 cm² to 5.29 cm².

• For pocket (undermining)—a long‐standing challenge because the wound margin cannot be visualised from the skin surface—the model achieved 94.1% concordance (weighted κ, 0.952) and halved the mean absolute error (14.02 cm² to 6.56 cm²), although a few outliers remained when shadows or ruler markings were misidentified.

1. Introduction

The increasing demand for cost containment and operational efficiency presents an urgent challenge for modern healthcare systems. Recently, the rapid advancement of artificial intelligence (AI) has gained attention as a potential solution to these challenges, with increasing applications in wound management, including pressure injuries [1, 2]. Regular pressure injury size measurement and undermining (also referred to as pockets or subcutaneous detachment beneath the wound edge) are essential for the objective evaluation of wound‐healing progress. Notably, automated quantification of these parameters from clinical photographs could help standardise wound assessment and reduce clinical workload.

We previously reported AI‐based studies on pressure injury assessment, including the optimization of image acquisition conditions and the development of models for depth and necrotic tissue evaluation [3, 4, 5]. Although several studies have proposed automatic quantification of ulcer size, previous research indicates that, in pressure injuries with undermining, the wound margin cannot be continuously visualised from the skin surface, making area estimation from images alone challenging [6, 7, 8]. Therefore, the development of algorithms capable of handling undermining accurately remains an unresolved issue.

Traditionally, polygonal contour annotation using LabelMe has been used for wound delineation. Nonetheless, LabelMe‐based annotation can lead to the overrecognition of non‐wound regions. Recently, object detection methods using You Only Look Once (YOLO) have shown promising results for wound localization [9]. Here, we developed an AI model that integrates LabelMe‐based annotation and YOLO detection to quantitatively assess pressure injury size and undermining. This study aimed to evaluate the performance of YOLO‐assisted lesion detection and validate the model's capability to assess undermining, a persistent challenge in pressure injury analysis.

2. Materials and Methods

2.1. Research Design

This was a retrospective, observational study. Photographs of the pressure injuries and corresponding wound measurements for size and pocket (undermining) were collected from electronic medical records. Using these data, we developed an AI model to recognise the pressure injury size and pocket extent from wound images and evaluated its performance.

2.2. Research Approval

This study was approved by the Institutional Ethics Committee of Tottori University Hospital (Approval No. 21A153). The study overview was posted on the hospital website and in the outpatient clinics to ensure participants' right to opt out.

2.3. Supervised Learning

A supervised learning approach was adopted for the model development. In supervised learning, annotated datasets are split into training and test sets; the AI model is trained on the training set and its performance is subsequently evaluated on the test set.

2.4. CasMatch

CasMatch (Bear Medic Co. Ltd., Japan) is a 10 × 10 mm colour calibration chart composed of nine colour tones, serving as both a chromatic and dimensional reference. When placed adjacent to the wound during image capture, the CasMatch enables post hoc measurement of wound size via image recognition [3].

2.5. DESIGN‐R

In this study, an AI model was developed based on the DESIGN‐R scale, a widely used pressure injury assessment tool developed by the Japanese Society of Pressure Ulcers in 2008 [10]. The DESIGN‐R evaluates seven parameters: depth (D), exudate (E), size (S), inflammation/infection (I), granulation (G), necrotic tissue (N) and pocket (P, undermining). This study focused on evaluating the S (size) and P (pocket) components.

The S component represents the surface area of tissue damage, calculated by multiplying the longest diameter (a) by the maximum perpendicular short diameter, measured in centimetres (Figure 1a). The size score is categorised into seven stages: s0 (none), s3 (< 4 cm²), s6 (≥ 4 and < 16 cm²), s8 (≥ 16 and < 36 cm²), s9 (≥ 36 and < 64 cm²), s12 (≥ 64 and < 100 cm²) and S15 (≥ 100 cm²). Mild stages are denoted by lowercase letters and severe stages are indicated by uppercase letters (Table 1).

FIGURE 1.

Measurement methods for the S (size) and P (pocket, undermining) components in the DESIGN‐R classification. (a) The area of skin damage (S) was calculated as S = a × b (cm²), where a represents the longest diameter and b indicates the maximum perpendicular short diameter of the skin damage. (b) The pocket area (P) was calculated as P = c × d – a × b (cm²), where c is the longest diameter including the pocket and d is the maximum perpendicular short diameter including the pocket.

TABLE 1.

DESIGN‐R scoring.

S (Size): the area of skin injury (length × width)
s	0	None	S	15	≥ 100 cm2
	3	< 4 cm2
	6	≥ 4 and < 16 cm2
	8	≥ 16 and < 36 cm2
	9	≥ 36 and < 64 cm2
	12	≥ 64 and < 100 cm2
P (Pocket, undermining): the area obtained by subtracting the ulcer from the entire affected area, including the pocket
p	0	None	P	6	< 4 cm2
				9	≥ 4 and < 16 cm2
				12	≥ 16 and < 36 cm2
				24	≥ 36 cm2

The P component represents the pocket area, defined as the subcutaneous layer beneath the wound edge. The total wound outline, including the pocket, was delineated and the longest and maximum perpendicular short diameters were measured (Figure 1b). The pocket area (P) was calculated by subtracting the skin‐injury area (S) from the total outline area. Pocket severity is classified into five stages: p0 (none), P6 (< 4 cm²), P9 (≥ 4 and < 16 cm²), P12 (≥ 16 and < 36 cm²) and P24 (≥ 36 cm²). Similar to S, lowercase letters denote milder stages and uppercase letters denote more severe stages (Table 1). The non‐sequential numbering of stages reflects their function as a weighted scoring system [10].

2.6. YOLO

YOLO is a one‐stage object detection algorithm that uses convolutional neural networks to extract image features and simultaneously estimate object location and class. It requires a sufficiently large training dataset for model construction and has previously been applied to the automatic detection of pressure injuries [9].

2.7. Subjects

Patients aged 20 years or older who were examined for pressure injuries at our hospital between January 2014 and December 2022 were included in this study. From electronic medical records and clinical databases, wound photographs and corresponding DESIGN‐R findings for the size (S) and pocket (P) components were retrospectively collected.

2.8. Dataset Labelling and Processing

Wound images were captured under nonstandardised conditions by multiple healthcare professionals. All images were stored in JPEG format and displayed in RGB colour mode. At our hospital, the pocket (P) area is routinely marked with black dotted lines for measurement; therefore, wound images containing pockets (P) included these markings. The DESIGN‐R scores for size (S) and pocket (P) recorded in the clinical charts were verified by the authors, who are plastic surgeons. A labelled dataset (all datasets) containing each wound image and the corresponding S and P scores was created. The dataset was first split into training and test sets at the image level, ensuring that no identical image appeared in both sets. Because the split was performed at the image level, different images from the same patient could be assigned to both sets. Data augmentation was then applied exclusively to the training set, performed by rotating each image by 10° intervals and increasing the dataset size by 36‐fold.

2.9. AI Model Building

Here, two models were developed to recognise the skin injury area (S) and pocket (P) areas from clinical photographs and to calculate the corresponding DESIGN‐R sizes (S) and pocket (P) areas and stages. The models used were the LabelMe‐Seeded Auto‐Recognition (LSAR) and YOLO‐gated LSAR (YGL) models.

The common procedure for both models is as follows. A colour calibration chart (CasMatch) was automatically recognised within each image. According to the DESIGN‐R criteria, the longest and maximum perpendicular short diameters of the skin injury area were measured to calculate the S‐area. The pocket (P) region, indicated by black dotted lines, was identified and the longest and shortest perpendicular diameters of the total wound area, including the pocket (P), were measured. Subsequently, the pocket (P) area was calculated as ‘(skin injury + pocket) area − S area’.

The specific structure and differences between the two models were as follows:

For the LSAR model, a subset of images is manually annotated in LabelMe using polygonal contours to create training labels. The model was then trained using these labelled data and the wound regions were automatically extracted from the unseen images.

The YGL model incorporates LabelMe and YOLO for combined segmentation and object detection. In the LSAR‐only approach, excessive regions are sometimes detected as part of the wound (S), leading to an overestimation of the wound size during area calculation. Accordingly, the YGL model employs YOLO‐based gating to localise the wound region, restricting the area analysed by the LSAR model and preventing the detection of unnecessary regions (Figure 2).

FIGURE 2.

Differences in annotation regions between the LSAR and YGL models. Yellow: Region detected using LabelMe only (LSAR model). Red: Region detected by YOLO. Blue: Region detected by LabelMe within the YOLO‐detected area (YGL model).

Both models were trained using the training dataset. During inference, S and P were calculated using the procedures described above. The primary difference between the two models lies in their region‐extraction methods: the LSAR model automatically extracts wound regions after training on LabelMe‐generated labels, whereas the YGL model uses YOLO to assist the LSAR model by defining the regions to be analysed.

2.10. Performance Evaluation and Statistical Analysis

The performances of the two models were evaluated by comparing the DESIGN‐R stages and areas calculated from clinical photographs with reference labels. The DESIGN‐R stages included S = {0, 3, 6, 8, 9, 12, 15} and P = {0, 6, 9, 12, 24}. For stage classification, the primary evaluation metrics were simple accuracy and the weighted Cohen's κ coefficient with quadratic weighting to account for ordinal relationships. For the area assessment (cm²), errors were treated as continuous variables and the mean absolute error (MAE) and median (interquartile range [IQR]) were used as summary statistics. Confidence intervals (CIs) for accuracy were estimated using the Clopper–Pearson exact method and those for weighted κ were derived by bootstrap resampling (5000 iterations, percentile method). Analyses of the wound area were conducted using valid cases, excluding those with missing data.

All analyses were performed on a workstation equipped with an Intel Core i9‐11900K CPU@3.50 GHz (16 cores), 64.0 GB memory and an NVIDIA GeForce RTX 3090 GPU. Data were aggregated using Microsoft Excel (Version 2509). Annotation was performed using LabelMe (version 5.1.1) and YOLO (Version 8). Data augmentation, CasMatch recognition and measurement of the skin injury and pocket (P) areas were performed using custom Python scripts (version 3.10.0). Statistical analyses were conducted using R (version 4.5.1) and RStudio (Version 2025.05.1 + 513).

3. Results

Overall, 1017 pressure injury images from 329 patients were collected and divided into training (n = 979) and test (n = 38) datasets. The training images comprised 882 from 265 patients for the S model, 116 from 28 patients for the P model and 979 from 311 patients for the YOLO model. The 38 test images were obtained from 11 patients (21 from 6 patients for the S model and 17 from 5 patients for the P model). For YOLO‐based wound region recognition, all 979 annotated training images were used as labelled data. The 38 test images, each containing CasMatch within the photographic field, were independent of the training dataset (Figure 3).

FIGURE 3.

Overview of model building and evaluation processes. LSAR, LabelMe‐Seeded Auto‐Recognition; YGL, YOLO‐gated LSAR; YOLO, You Only Look Once.

For the classification concordance of DESIGN‐R stages, the LSAR model demonstrated an accuracy of 61.9% and a weighted κ of 0.492 (95% CI, 0.188–0.667) for S (size). The YGL model demonstrated markedly higher performance, with an accuracy of 85.7% and a weighted κ of 0.969 (95% CI, 0.625–1.000). For P (pocket) classification (n = 17 for each model), the LSAR model achieved an accuracy of 88.2% and a weighted κ of 0.783 (95% CI, 0.361–1.000), whereas the YGL model demonstrated an accuracy of 94.1% and a weighted κ of 0.952 (95% CI, 0.816–1.000). Both models exhibited good agreement; nevertheless, the YGL model demonstrated superior concordance (Table 2).

TABLE 2.

Stage evaluation.

Task	Model	Accuracy (%)	Weighted κ	95% CI
S	LSAR	61.9	0.492	0.188–0.667
S	YGL	85.7	0.969	0.625–1.000
P	LSAR	88.2	0.783	0.361–1.000
P	YGL	94.1	0.952	0.816–1.000

Abbreviations: CI, confidence interval; LSAR, LabelMe‐Seeded Auto‐Recognition; P, pocket; S, size; YGL, YOLO‐gated LSAR.

For the area error analysis of S, the LSAR model demonstrated an MAE of 15.22 cm² (95% CI, 1.20–42.43) and a median (IQR) of 1.82 cm² (0.48–3.15). The YGL model exhibited smaller errors, with an MAE of 5.29 cm² (95% CI, 1.09–12.98) and a median (IQR) of 1.10 cm² (0.59–2.69). For P, the LSAR model yielded an MAE of 14.02 cm² (95% CI, 5.80–24.40) and a median (IQR) of 6.62 cm² (2.25–12.98), whereas the YGL model yielded an MAE of 6.56 cm² (95% CI, 3.32–10.46) and a median (IQR) of 3.74 cm² (1.16–7.92), smaller than those of the LSAR model for both metrics (Table 3).

TABLE 3.

Area error.

Task	Model	MAE (95% CI)	Median MAE (IQR)
S	LSAR	15.22 (1.20–42.43)	1.82 (0.48–3.15)
S	YGL	5.29 (1.09–12.98)	1.10 (0.59–2.69)
P	LSAR	14.02 (5.80–24.40)	6.62 (2.25–12.98)
P	YGL	6.56 (3.32–10.46)	3.74 (1.16–7.92)

Abbreviations: CI, confidence interval; IQR, interquartile range; LSAR, LabelMe‐Seeded Auto‐Recognition; MAE, mean absolute error; P, pocket; S, size; YGL, YOLO‐gated LSAR.

4. Discussion

This study compared two models for automated calculation of the DESIGN‐R parameters S and P, along with the corresponding wound area: the LSAR and YGL models.

The YGL model consistently demonstrated a higher concordance in stage classification. For S, the YGL model achieved higher accuracy and weighted Cohen's κ values and similar trends were observed for P. Because the weighted κ (squared) reflects the ordinal nature of the data and penalises larger discrepancies more severely, these results indicate that the YGL model successfully minimised major misclassifications and limited discrepancies mainly to adjacent stages.

In the area‐error analysis, the YGL model outperformed the LSAR model for both S and P, although a few outliers remained for P. For S, the YGL model demonstrated smaller errors for both the MAE and median (IQR). For P, the YGL model also showed a smaller MAE and median (IQR) than the LSAR model, although a few large outliers remained. The P parameter represents undermining beneath the wound edge, which cannot be visualised directly from the surface. Both models assume the presence of black dotted markings to delineate the undermining area. Nonetheless, in some cases, shadows or black‐scale markings from rulers in the image were misidentified as pockets (Figure 4). Such misrecognition likely leads to an exaggerated area estimation, resulting in squared‐error amplification and a few large measurement outliers.

FIGURE 4.

Example of a photograph in which the black scale markings were misrecognised as the black dotted lines indicating the pocket (P) area.

In future developments, the accuracy of the wound size and pocket (P) measurement is expected to improve further by enabling the AI model to automatically recognise and display the wound area within the application, thereby allowing users to verify and adjust the detected region in real time.

4.1. Limitations

This study had some limitations. First, the sample size was relatively small and only 38 test images were included. This was due to the limited number of photographs that appropriately incorporated CasMatch and were therefore suitable for quantitative evaluation. Accumulating a larger sample size remains a task for future work. Second, patient‐level overlap exists between the training and test sets for the P model. Although no image‐level duplication remained after the retraining described above, all five patients contributing to the 17 test images also had other photographs among the 116 training images (40 images in total). This reflects the limited number of patients whose photographs met the inclusion criteria for quantitative P analysis. On review of these images, the pocket size had changed over the course of wound healing and the thickness and spacing of the black dotted markings also varied among images; these images were therefore not repeated records of an identical finding. Nevertheless, patient‐specific characteristics such as body habitus, skin tone, wound location and photographic environment may have been shared, so the reported P‐model performance may not fully reflect generalization to unseen patients and should be interpreted with this constraint in mind. Confirming patient‐independent generalization will require larger and more diverse patient datasets in future work. Third, the method depends on CasMatch. In the present method, stage classification is derived from the calculated area, which depends on CasMatch for metric calibration, so CasMatch is required even for stage‐level assessment. If wound size could instead be estimated using intrinsic anatomical references, such as representative skin wrinkles, CasMatch might become unnecessary and the test set could be expanded; this remains an expectation for future technical development and research rather than a current capability. Fourth, pocket (P) area measurement depends on the black dotted markings. However, because the superficial skin overlying a pocket may or may not exhibit visible colour changes, some form of marking is considered essential for accurate delineation. Automating the marking itself—for example, by detecting pockets through skin surface discoloration or texture analysis—or assessing pockets without any markings would further reduce the clinical workload; however, assessment without black dot markings remains difficult at present and such automation has not yet been achieved, representing an expectation for future research. Fifth, the model evaluates only the S and P components of DESIGN‐R. Of the remaining parameters, automated assessment of depth (D) and necrotic tissue (N) has been reported previously by our group [4, 5], while exudate (E), inflammation/infection (I) and granulation tissue (G) remain to be addressed in future work. Sixth, all the patients included in this study were Japanese and the dataset lacked samples from individuals with different skin tones. Therefore, the performance of this model has not yet been evaluated in non‐Asian populations. Future studies with larger and more ethnically diverse datasets are warranted to validate and generalise these findings.

In conclusion, combining LabelMe with YOLO improved wound recognition accuracy compared with LabelMe alone. Furthermore, the long‐standing challenge of pocket (P) evaluation was addressed by recognizing marked areas. This approach represents a useful step towards the practical implementation of automated measurements in clinical settings and it is expected to contribute to reducing the clinical workload.

Funding

This research was supported by the Research Support Program of Tottori University Hospital and Japan Society for the Promotion of Science (JSPS) KAKENHI (Grant Number 24K12824).

Disclosure

The authors collaborated with Focus Systems Corporation in developing the AI model described in this study. However, there was no financial support, employment or personal relationship between the authors and the company.

Ethics Statement

This retrospective observational study was approved by the Institutional Ethics Committee of Tottori University Hospital (Approval No. 21A153) and conducted in accordance with the Declaration of Helsinki.

Consent

Given the retrospective design, the requirement for individual informed consent was waived and participants were given the opportunity to opt out via a public notice.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.