In this stage, the primary objective was to identify relevant studies on DFU wound healing prediction using imaging techniques. PRISMA guidelines were used to identify journal articles published in the last decade, (July 2014 to July 2024) that reported wound healing prediction algorithms utilizing imaging techniques. Publications were obtained from five online databases that included engineering and clinical journals (Web of Science, Scopus, CINAHL, Embase, and PubMed) using an advanced search process. The following search terms, developed in consultation with the university librarian, were used; (“diabetic foot” OR “diabetic foot ulcers” OR “diabetic-related foot ulcers” OR “neuropathic diabetic foot ulcer”) AND (“imaging technique” OR “imaging system” OR “hyperspectral imaging” OR “thermal imaging” OR “photographic imaging”) AND (“prediction” OR “predicting” OR “predict”) AND (“healing status” OR “healing time” OR “healing”). Only English-language journal articles were included; non-English studies were excluded due to resource limitations for accurate translation. Preprints and unpublished data were not considered, as filters were applied during the initial database search to include only peer-reviewed, published journal articles. This approach ensured that all included studies went through formal scientific review, reduced the risk of bias associated with unverified or incomplete findings.

After collecting the studies, the next stage involved screening the articles. A PRISMA flowchart is shown in Figure 2, which was used to identify the articles. EndNote (Clarivate, UK) was used to identify and remove duplicates. After the removal of 104 duplicates, the searching process identified 2,833 publications to be considered for the screening step.

The screening process was conducted in two steps. In the first step, the title and abstract of each study were reviewed by two reviewers (NNS, NDP) to generate a list of relevant studies based on the inclusion and exclusion criteria. Grey literature, including posters, clinical notes, meeting reports, case reports, and news articles (n = 457), was excluded to maintain consistency in quality assessment and ensure inclusion of studies that went through the peer review process. The exclusion criteria included; studies that did not involve human subjects (n = 213); studies that not related to DFU healing prediction review articles (n = 76); studies not relevant to DFU (n = 1,309); those focused on diabetic foot infections or osteomyelitis (n = 66); studies that did not utilize imaging techniques (n = 626); those unrelated to DFU wound healing or healing prediction (n = 40); that focused on wound area segmentation or localization only (n = 11).

If the information from the title and abstract is insufficient to decide on the inclusion or exclusion of a study, the full text was reviewed. The second step involved a full-text review by the two reviewers (NNS, NDP) to obtain the final list of studies for data extraction. The inclusion criteria were: (1) original articles focused on predicting the healing status of diabetic foot ulcers (DFU) using imaging techniques (n = 17) and (2) review articles discussing imaging methods relevant to DFU wound healing (n = 5). Studies with insufficient information on outcomes or imaging methodology were excluded (n = 13). Any conflicts between the two reviewers were resolved by input from a third reviewer (QN).

After the two-stage screening process, as shown in Figure 2, 22 publications (17 journal articles and 5 systematic reviews) were identified. Nine articles were from clinical journals, eight were from engineering journals, and five were review articles, all of which were from the clinical domain. The CASP (Critical Appraisal Skills Programme) checklist for appraising quantitative studies was used for a quality checklist of included articles as shown in Table 3. Fifteen articles met the criteria for high-quality studies, while two met fewer criteria. No studies were removed based on quality because all were regarded as giving valuable insight.

While only 17 studies were included out of 2,937 screened, this targeted selection allowed for a more in-depth evaluation of advanced image modalities. This focused inclusion ensured that each selected study contributed valuable insights into the prediction of DFU wound healing using non-invasive imaging techniques. This approach aligns with the PRISMA guidelines, which emphasize the importance of methodological rigor in systematic reviews, thus enhancing the reliability and depth of our findings.

From the review, we identified four main categories of wound healing prediction using non-invasive imaging techniques: (a) 3D imaging (25, 26), (b) visual (RGB) imaging (27–32), (c) hyperspectral imaging (33–38), and (d) thermal imaging (39–41). While 3D imaging was not included in the search keywords, it appeared in the results and hence has been included in this study due to the visibility of this device for use in clinical, community, and home care services.

In this stage, valuable data were extracted from the selected articles. The prediction algorithms used in each non-invasive imaging technique were identified in Section 3, the characteristics for wound healing were assessed in Section 4, and the prediction models were described in Section 5. Additionally, the strengths, weaknesses, and technological readiness levels of each imaging technique were evaluated in Section 6.

The final stage involved summarizing the findings from the data derivation process. Key insights were drawn, including future directions for research, open research issues, and the barriers that were faced in clinical practice (Sections 7 and 8). This stage provided a comprehensive summary of findings and future investigations of DFU wound healing management, especially for wound healing prediction using non-invasive imaging techniques.

Measuring the volume of the wound, which includes the depth of the wound has been proposed for better evaluation of the wounds. Two studies reported the use of 3D imaging techniques in DFUs (25, 26). This prediction technique typically follows a series of steps to evaluate wound healing, beginning with capturing the wound image using a specialized 3D camera system. One study involved an adhesive optical target for calibration (26).

Multiple images were captured from multiple angles after wound debridement, and these images were then combined into a single 3D model by the software. The clinicians then manually traced the wound area on the monitor for precise measurements to calculate various parameters such as wound area, volume, and depth. The wound measurements were compared from baseline to predict wound healing progression.

These measurements were analyzed using statistical models such as linear regression, which correlates the changes in wound dimensions with time to healing. Prognostic indicators, such as the healing slope and regression analysis, determined whether the wound was healing as expected or stagnating. One study (25) used measurement from 2D imaging as a comparison, while another (26) used the complete closure of the wound (100% skin regrowth) as a reference.

Visible light photographic imaging, generally performed using RGB (Red, Green, Blue) imaging, is one of the imaging techniques that closely resembles traditional assessment in clinical practice due to its focus on measuring physical appearance such as length, width, depth, edge, and peri-wound skin to monitor the wound healing progression (42, 43).

In the list of included articles, six studies focused on using photographic imaging techniques to predict wound healing progression in DFU patients (27–32). The algorithms for predicting healing in DFUs mostly involve four main processes: image acquisition, preprocessing, segmentation, and classification. The algorithm steps used for predicting the healing progression are described below.

Hyperspectral imaging (HSI) is a technique that determines the high-resolution spectrum of each pixel across the electromagnetic spectrum of interest. Six studies used HSI to predict the progression of DFU healing (33–38). As the first step, the eligible patients were properly screened based on inclusion criteria (35, 37, 38), for example, by abstaining from smoking or caffeine (33). To reduce noise and normalize attenuation values, the calibration of a push-broom camera across a wavelength range of 430–750 nm was applied in the study conducted by Yang et al. (34). Hyperspectral imaging data, corrected for light scattering (33), was processed to calculate tissue oxygenation levels (33), with skin temperature and the ankle-brachial pressure index (ABPI) recorded as baseline measurements (33). In another study, baseline assessments included transcutaneous oxygen pressure (TcpO2), ankle-brachial and toe-brachial indexes, and HSI, which evaluated tissue parameters such as oxygen saturation, hemoglobin index, near-infrared perfusion, and water content (37). The HSI was also used to quantify peri-wound oxyhemoglobin (OxyHb), deoxyhemoglobin (DeoxyHb), and oxygen saturation (O2Sat) in two studies (36, 38). Photonics-based algorithms were used in one study (36) and included the combination of HSI and thermal imaging applied to capture data on biomarkers including OxyHb, DeoxyHb as well as foot temperature patterns to compute the oxygenation metrics Oxygen saturation (SpO2) and Tissue oxygen saturation (StO2).

Monitoring of wound healing progression was carried out at 12 weeks (33, 38) and 24 weeks (33, 37), while other studies conducted monitoring over a 3-week period to assess changes in wound size, perfusion [e.g., transcutaneous oxygen levels (36), hyperspectral imaging], and cytokine levels (35). Using these features, wounds were classified as “healers” or “non-healers,” and a predictive model was created to estimate the likelihood of wound healing. In another study (37) healing was defined as complete epithelization without drainage sustained for at least 10 days within 24 weeks.

Thermal imaging records the infrared or near-infrared radiation of the object to assess its thermal condition. It is a non-invasive technique that detects temperature variations on the surface of an object. In the prediction of wound healing of DFU, in the study by (28) thermal images were segmented into isothermal patches to analyze the wound boundaries and the areas according to the temperature distribution (39). The assessment parameters include the temperature of the wound bed, the area of the isothermal patch on the wound bed, the area of the isothermal patch surrounding the wound and the number of isolated isothermal patches within the wound region. The physical area of the wound bed was determined from color images (39).

In another study (40), the combination of thermal imaging and computerized planimetry was used to first monitor the temperature change indicative of metabolic activity and inflammation, and second to measure wound surface area and wound perimeter. This study integrated these measurements to evaluate the effect of Hyperbaric Oxygen Therapy (HBOT) (40).

Deep learning-based methodology was used in the combination of thermal and visible imaging (41). Experts annotated the fused image as an input to the Mask R-CNN model for precise wound segmentation. Weekly images were analyzed to calculate metrics (39–41) including ulcer-to-foot area ratio and mean temperature distribution. These metrics were correlated with clinician-provided ground truth data to track the healing progression.

Evaluating the physiological aspects of wounds by hyperspectral imaging is essential for forecasting healing trajectory. This assessment concentrates on internal tissue properties and physiological states. Wound healing is a complex and dynamic process that involves four connected stages: blood clotting (hemostasis), inflammatory response, tissue growth (proliferation), and tissue repair (44–46). A proper understanding of the four stages of wound healing is essential for the physiological assessment of wounds (47, 48).

Geometric wound assessment is an important predictor of healing status for DFUs, including the shape, wound size, perimeter, surface area, and volume measurements (25, 26, 43, 54, 55). These quantitative metrics provide objective and reproducible data that can help clinicians monitor wound progression over time, evaluate treatment efficacy, and predict healing outcomes (56, 57). The linear nature of wound margin advance (WMA) over time, calculated as the change in wound area divided by wound perimeter, has been shown to be an effective predictor of healing in DFUs (56). Medical professionals can use this linear pattern to estimate the total healing time based on assessments undertaken at 4–5 weekly intervals. This capability makes it a valuable asset for timely intervention and fine-tuning treatment approaches (56). Studies have shown that a 40% reduction in wound surface area within the first two to three weeks of treatment predicts healing within 12–24 weeks (57), though there is scant evidence to support per cent area reduction in isolation as a surrogate for complete DFU healing in routine clinical practice (8).

An analysis of surface features, textures, and segmentation may be used for understanding wound structure and valuable predictor of healing status in DFUs (58). They provide objective, quantifiable data on wound progression, including color classification (27, 30, 31), granulation area (RYB distinction) (29), texture feature (30), siamese neural network features (32), Mask R-CNN for wound segmentation (41), changes in ulcer perimeter and surface (40) and scattering analysis for detecting wound bed and surrounding tissue (33).

Statistical methods were selected based on the assumption that the healing of DFUs follows a linear pattern in wound area reduction (25) (26, 54, 59, 60). Jorgensen et al. (25) compared the changes in 2D and 3D area measurements using statistical approaches. Linear regressions were used to estimate the mean changes from each patient, with the negative values as healing (decreasing in size) and positive values as nonhealing (increasing in size). The paired Wilcoxon rank test was used to test the zero median difference of the slope between 3D and 2D area measurements. The Bonett-Price approach was then used to provide the confidence interval from that median difference and the Spearman correlation coefficient was used to calculate the correlations between changes in 2D and 3D area measurements (25). Data from 150 wounds revealed that while 2D and 3D area measurements changed in the same direction, the magnitude of changes in 3D measurements was consistently greater than those observed in 2D measurements (25).

Malone et al. (26) investigated the relationship between mean wound healing measurement variables obtained from five 3D wound measurements from 21 participants using linear regression and Pearson's correlation coefficient. The performance analysis metrics were the linear healing slope and statistical significance, with a p-value of 0.0001 and an R-value greater than 0.70, respectively.

Kartika et al. (29) proposed an alternative statistical analysis approach for assessing the wound healing process in DFUs. The researchers quantified wound area, granulation area, granulation index, and respective temporal changes using ImageJ software, a free web-based software that allows easy editing, display, and analysis of images. The statistical analysis was subsequently performed to evaluate the healing progression using SPSS version 20. There were statistically significant differences in granulation index on day 3, day 7, and day 14 (p < 0.05).

Yang et al. (34) used HSI to capture reflectance spectra from tissue and analyzed them to predict wound healing by examining SpO2 levels. The study compared two approaches: SpO2 measurement and Principal Component Analysis (PCA). The PCA, which reduces dimensionality by identifying principal components that capture relevant features of the data, outperformed SpO2 in predicting the healing by the 12th week. While both methods showed high specificity, PCA achieved higher sensitivity (87.5% vs. 50%) and better overall predictive accuracy, with a positive predictive value of 91.3%. The study demonstrated that PCA, using HSI data, was more effective than traditional SpO2 analysis in predicting the healing of DFUs.

In 2022, Lavery et al. (35) evaluated HSI from 23 wounds to compare continuous variables and healer vs. non-healer status using the paired t-test and independent t-test, respectively. Predictive features were derived from baseline data and weekly trends, focusing on achieving a ≥50% reduction in wound area (healer), improvements in perfusion, and significant changes in cytokine levels.

Kounas et al. (38) employed the two-sample t-test to distinguish between groups, a binary logistic regression analysis comparing healers and non-healers based on oxy-Hb levels. There was a negative correlation between oxyHb levels at Visit 1 and the percentage of ulcer size reduction from Visit 1 to Visit 4 (r = −0.46, p = 0.02), and between oxyHb levels at Visit 2 and the percentage of ulcer size reduction from Visit 2 to Visit 4 (r = −0.65, p = 0.001).

Lopez-Moral et al. (37) applied logistic regression to predict DFU healing status using clinical and imaging-derived features. Key predictors included Transcutaneous oximetry (TcpO2), HSI parameters i.e., Skeletal Muscle Oxygen Saturation (StO2), Thermal Hyperspectral Imager (THI) parameters, Near Infrared Perfusion Index (NIR-PI), Tissue water index (TWI), and vascular indices [i.e., ankle-brachial index (ABI) and toe-brachial index (TBI)]. P-values below 0.05 were considered statistically significant, with a 95% confidence interval (CI). Model performance was assessed via receiver operating characteristic (ROC) curve analysis, identifying optimal cut-off values for each predictor to maximize sensitivity and specificity. TcpO2 demonstrated the highest diagnostic accuracy (AUC = 0.989, p-value = 0.005), followed by StO2 (93% sensitivity and 71% specificity), establishing it as the strongest independent predictor of healing.

In a thermal imaging study, Aliahmad et al. (39) employed the nonparametric Kruskal–Wallis test to investigate the relationship between the healing status of ulcers at week 4 and the ratio of DFU parameters measured in the first two weeks of ulceration (week 2 to week 1 ratio) from the isothermal areas. They defined a timely healing ulcer trajectory as showing more than a 50% reduction in area by week 4, an indirect measure routinely used.

A study conducted by Glik et al. (40) explored the impact of Hyperbaric Oxygen Therapy (HBOT) on wound healing using a combination of planimetry and thermal imaging. Statistical analysis was based on the demographics, age of the treatment group, and the median values and ranges of the collected data. To evaluate relationships between variables, Spearman's rank correlation coefficient was calculated. For normally distributed data, paired-sample t-tests were employed, while Mann–Whitney U or Wilcoxon tests were used for data that did not follow a normal distribution. Levene's test was utilized to assess data normality. When comparing more than two groups, analysis of variance (ANOVA) was applied, with a significance threshold set at p < 0.05.

Recent advancements in machine learning and deep learning have significantly improved DFU monitoring and healing prediction. Various models have been developed to automate ulcer assessment, optimize treatment strategies, and enhance clinical decision-making. These approaches incorporate techniques like Random Forests, Support Vector Machines, Artificial Neural Networks, Siamese Neural Networks, and Convolutional Neural Networks, each utilizing different imaging modalities and features to provide accurate, non-invasive evaluations of wound healing.

Kim et al. (30) proposed the two machine learning models to predict the results of wound healing: Random Forests (RF) and Support Vector Machines (SVM) with RBF kernel. A total of 208 samples were used, with 25% allocated for testing and 75% used for 3-fold cross-validation training. A randomized grid search with three-fold cross-validation was used to optimize the models on the training set. Two thousand combinations of hyperparameters, such as bootstrapping, tree selection criteria, maximum depth, minimum samples per leaf, minimum samples to split a node, and number of PCA components, were tested for RF.

The PCA components, gamma (kernel coefficient), and C (error term penalty) were among the 2028 combinations tested for SVM. The optimal hyperparameters were chosen for each model from 300 randomly selected combinations to increase efficiency. The area under the receiver operating characteristic (AUROC) curve was used to assess the models' performance. To ascertain feature importance for prediction, a different RF model was trained on clinical and handcrafted image features without PCA. When trained with hand-crafted imaging features alone, both the RF and SVM models performed better than when trained with clinical or deep learning-based features alone (P < 0.05).

Motta et al. (31) developed an Artificial Neural Network (ANN) for classifying wound tissue into granulated, fibrin, and necrotic categories. The ulcer area was manually selected by a professional expert for use during training. A multilayer perceptron feedforward network with ten hidden neurons was employed. The Healing Ulcer Index (HUI), which quantified treatment outcomes, was designed for long-term comparison and to aid in monitoring patients at treatment centers. These indexes were based on changes in the wound area and variations in the proportions of different ulcer tissues.

Toofanee et al. (32) presented a novel framework utilizing Siamese Neural Networks (SNN) to monitor and assess the progression of DFU healing. The process began with data acquisition and preprocessing, including resizing and augmentation to address class imbalances. The framework trained an SNN using EfficientNetV2S and Vision Image Transformers to extract features and implement similarity learning, comparing new DFU images against class anchors—average feature representations for “none,” “infection,” “ischemia,” “both,” and “healthy” categories. During the evaluation, baseline images were compared with subsequent ones, generating radar charts and similarity scores to track healing progress and guide clinical decision-making. The model achieved a Macro F1-score of 0.6455, outperforming others in DFU classification tasks, and was validated by the end user, showing its utility for tracking DFU evolution.

Doulamis et al. (36) proposed a photonics-based device for DFU monitoring that integrates HSI and thermal imaging. This study has used the deep learning framework, including convolutional neural networks (CNNs), to implement the automatic segmentation and classification of pathological regions. The system supports dual configurations: a low-cost in-home version equipped with RGB and low-cost HSI sensors for patient self-monitoring, and a professional edition incorporating high-resolution HSI, near-infrared (NIR) fluorescence, and thermal sensors for clinical use. Data acquisition, processing, and interpretation were managed through an embedded software platform that integrates sensor input with machine learning models to deliver real-time, data-driven diagnostic support.

Sharma et al. (41) employed a deep learning-based approach using Mask R-CNN to evaluate DFU healing. The methodology integrated thermal and visual images through an HSV-based image fusion technique. Mask R-CNN, an instance segmentation model, accurately identifies and delineates the ulcer region, using a Region Proposal Network (RPN) and a refined ROI-Align operation. The model outputs both mask and class predictions independently, optimizing segmentation accuracy. This automated system correlates closely with clinician measurements (92.5% agreement).

The application of machine learning/deep learning in DFU imaging demonstrated the variety of algorithms for specific tasks such as classification and segmentation. Traditional machine learning models like Random Forests and Support Vector Machines used handcrafted features and improved their performance by tuning settings (e.g., randomized grid search) and reducing feature numbers (e.g., PCA). Deep learning approaches, including CNN, SNN, and Mask R-CNN learned features automatically from data, and used techniques like data augmentation to handle small datasets. The performance was measured using AUROC and F1-score, and the results were compared with expert labels as ground truth.

The lack of accessible wound images restricts the amount of data available for training and validating predictive models. As can be seen in Table 6, the sample sizes available in most of the included papers were relatively small, often around twenty, which may not be sufficient for machine learning approaches. Therefore, it is understandable that most of the predictions or measurements of healing were processed statistically. However, large and accessible datasets would be extremely valuable for improving the accuracy of predicting wound healing status in DFUs. Researchers may struggle to create models that generalize well across different populations and clinical scenarios in the absence of a diverse and comprehensive dataset.

In the clinical environment, efficient computational methods are needed to obtain quick and real-time results, including automated wound segmentation. Manual segmentation makes it unsuitable for the device to be used in the clinical setting, delays the process of analysis, and leads to data without clear standards, as different individuals might produce varying tracing results (25, 30).

Technical limitations must be considered before the device can be used in a wide range of clinical settings. Some studies have used complicated tools for data collection that are not portable or feasible for routine use in primary care settings (25, 27, 34, 41). Expensive imaging technologies such as HIS (25, 37) can also prevent many healthcare providers from using advanced tools to assess DFUs, especially since these devices are intended for use not only in clinical environments but also in home care settings.

Currently, there is a lack of large-scale datasets for DFU images. This represents a bottleneck in advancing predictive models for healing. Future efforts should focus on developing centralized, open-access repositories of annotated wound images that are collected from diverse populations and clinical settings (61–63). These repositories would enable researchers to train and test predictive models on a broader set of data, significantly improving the generalizability and robustness (64).

Techniques like synthetic data generation using methods such as generative adversarial networks (GANs), could also be explored to augment existing datasets and overcome data bias. Synthetic wound images have shown promise in enhancing dataset diversity, particularly in addressing class imbalances and rare scenarios (24, 65).

Combining wound imaging with complementary information such as patient demographics, comorbidities, biochemical markers, and treatment history could provide a more holistic understanding of the healing process. Multimodal approaches have been shown to improve model performance in predicting the healing of DFU (66) and chronic wounds (67). This integration would require collaborative efforts across disciplines to standardize data collection protocols and ensure interoperability between datasets.

Standard protocol in non-invasive image technique to predict healing status of DFU is essential to ensure consistent, reliable, and reproducible assessment. Consistent procedures for image acquisition such as lighting, positioning, calibration, environmental/camera control are critical to ensure reliable data across time and clinical settings. Protocol standardization can facilitate clinical implementation by making imaging more interpretable for healthcare providers.

Machine learning methods, especially convolutional neural networks (CNNs), such as U-Net and its variants, have proven highly effective in medical image segmentation and could be adapted for use in predicting DFU healing (68, 69). To enhance clinical adoption, these methods must be computationally efficient, capable of running on standard hardware, and robust to variations in wound appearance caused by lighting, skin tone, and ulcer severity.

In addition, the implementation of explainable AI (XAI) methods in segmentation tools would enhance the transparency and acceptance among clinicians (70). Providing visual overlays or confidence maps could help clinicians validate the automated segmentation results, ensuring the reliability in decision-making processes.

To validate the predictive utility of imaging and AI models, multicenter prospective clinical trials are needed. These should evaluate whether the early prediction of healing status leads to improved patient outcomes, such as earlier intervention or reduced amputation. Trials should also assess how imaging tools perform across different care settings, such as hospitals, community, and home care services.

Future research should focus on developing portable, cost-effective devices that maintain high diagnostic accuracy. Advances in sensor miniaturization, wearable technology, and smartphone-based imaging platforms could provide practical solutions, enabling real-time wound assessment and monitoring in diverse care environments (71, 72). In addition, the integration of user-centered design principles is critical to ensure that these tools are intuitive for non-specialist users, such as general practitioners and caregivers (73, 74). Efforts should also be made to reduce the computational requirements of data analysis algorithms, enabling deployment on low-power devices.

Combining multiple image techniques into an integrated framework provides a more comprehensive view of wound healing status by capturing physical, physiological, and biochemical information. These can improve diagnostic capabilities beyond the limitations of individual techniques. Future direction should focus on developing tools that can quickly and accurately analyze the combined imaging features, which make it easier for clinicians to assess wounds and make treatment decisions.