Developing an AI-powered wound assessment tool: a methodological approach to data collection and model optimization

Abstract

Background

Chronic wounds (CWs) represent a significant and growing challenge in healthcare due to their prolonged healing times, complex management, and associated costs. Inadequate wound assessment by healthcare professionals (HCPs), often due to limited training and high clinical workload, contributes to suboptimal treatment and increased risk of complications. This study aimed to develop an artificial intelligence (AI)-powered wound assessment tool, integrated into a mobile application, to support HCPs in diagnosis, monitoring, and clinical decision-making.

Methods

A multicenter observational study was conducted across three healthcare institutions in Western Switzerland. Researchers compiled a hybrid dataset of approximately 4,000 wound images through both retrospective extraction from clinical records and prospective collection using a standardized mobile application. The prospective data included high-resolution images, short videos, and 3D scans, along with structured clinical metadata. Retrospective data were anonymized and manually annotated by wound care experts. All images were labeled for wound segmentation and tissue classification to train and validate deep learning models.

Results

The resulting dataset represented a broad spectrum of wound types (acute and chronic), anatomical locations, skin tones, and healing stages. The AI-based wound segmentation model, developed using the Deeplabv3 + architecture with a ResNet50 backbone, achieved a DICE score of 92% and an Intersection-over-Union (IOU) score of 85%. Tissue classification yielded a preliminary mean DICE score of 78%, although accuracy varied across tissue types, especially fibrin and necrosis. The models were optimized for mobile implementation through quantization, achieving real-time inference with an average processing time of 0.3 seconds and only a 0.3% performance reduction. The dual approach to data collection, prospective and retrospective—ensured both image standardization and real-world variability, enhancing the model’s generalizability.

Conclusions

This study laid the foundation for an AI-driven digital tool to assist clinical wound assessment and education. The integration of robust datasets and AI models demonstrated the potential to improve diagnostic precision, support personalized care, and reduce wound-related healthcare costs. Although challenges remained, particularly in tissue classification, this work highlighted the promise of AI in transforming wound care and advancing clinical training.

Trial registration

Not applicable.

Key Points

This study aims to develop an AI-powered wound assessment tool that supports clinical decision-making and enhances wound care quality through accurate tissue segmentation, wound monitoring, and healing prediction.

A large and diverse wound image database, including over 4000 wound assessments—will be created to train and validate AI models, ensuring clinical relevance across various wound types and patient profiles.

This innovation addresses a critical gap in wound care expertise, aiming to reduce diagnostic variability, support guideline adherence, and ultimately improve patient outcomes while lowering healthcare costs.

Background

The management of wounds, particularly chronic wounds (CWs), remains a significant challenge in modern healthcare due to their complex pathophysiology, slow healing progression, and high resource utilization [1]. CWs, including diabetic foot ulcers, venous leg ulcers, and pressure ulcers, are often associated with comorbid conditions such as diabetes mellitus, cardiovascular disease, autoimmune disorders, and peripheral artery disease, which further complicate their healing process [2]. The global prevalence of CWs is substantial, affecting approximately 2.2 per 1,000 people living with CWs at any given time [3]. The economic burden is also considerable, for example, the annual cost of managing CWs in the United Kingdom is estimated at £8.3 billion [4].

Wound care is gaining importance due to the aging global population, as older adults are more prone to developing chronic wounds due to age-related physiological changes such as reduced skin elasticity, vascular impairments, and a higher prevalence of chronic illnesses [5]. Assessing wound healing progress is crucial yet remains a time-consuming and often subjective task. Key factors in wound assessment include wound size, tissue composition (granulation, epithelialization, necrotic tissue), infection status, and treatment response [6, 7]. However, current wound measurement techniques, such as manual rulers, acetate tracings, and planimetry, present limitations in terms of accuracy, efficiency, and standardization [8, 9]. Ruler-based measurements, while widely used, have been shown to overestimate wound size by as much as 40%, and acetate tracings, though more precise, introduce additional risks such as contamination and discomfort for patients [7, 10]. Conventional non-automatized digital imaging methods, including computer-assisted planimetry and image-based analysis, have demonstrated improved accuracy but still require manual input, making them time-intensive [7].

Recent advances in artificial intelligence (AI) and machine learning (ML) have demonstrated significant potential to revolutionize wound assessment and management [11]. AI-based tools leverage machine learning (ML) and deep learning (DL) algorithms to analyze wound images and automate the evaluation process. These technologies have already been successfully integrated into medical imaging, including radiology, dermatology, and pathology, enhancing diagnostic accuracy and efficiency [12].

Within wound care, AI has shown promising results in improving wound size measurements, tissue classification, and healing outcome predictions. AI-driven models have been developed to automate wound segmentation and granulation, reducing inter-clinician variability and improving accuracy [13, 14]. Research has demonstrated that AI-based wound assessment algorithms can perform comparably to expert human annotations, with statistically similar error distributions between AI and human tracings [7, 15]. Moreover, AI-assisted wound monitoring has the potential to optimize treatment strategies, reduce clinical workload, and enhance patient outcomes [16].

Despite the advancements in AI-driven wound care solutions, their clinical applicability depends on high-quality, diverse datasets for model training and validation. Although such dataset exists, they are often limited to specific types of wounds (e.g. diabetic foot ulcers), focused on specific tasks (e.g. wound segmentation), or contain a relatively limited number of samples [17, 18]. In contrast, AI models require large, representative datasets to ensure robustness and generalizability across different wound types, tasks, patient populations, and clinical settings [6, 19]. However, collecting such data presents multiple challenges, including patient privacy concerns, variations in image acquisition techniques, and standardization of wound annotations.

This study aims to detail the methodology and challenges associated with data collection for the development of an AI-based wound assessment tool. Specifically, we collaborated with a Swiss startup to create a mobile application for clinical decision-making in wound care. In a future project, we also plan to develop a full serious game (SG) using the collected data. We outline the steps involved in data acquisition, discuss the barriers encountered, and highlight key facilitators that enabled efficient data collection.

Method

Developing an AI-powered wound assessment tool required the collection and processing of a diverse set of wound images along with detailed clinical data. To ensure a robust and representative dataset, wound images were obtained from multiple healthcare institutions using both retrospective and prospective approaches. The combination of these methods allowed for the integration of a standardized imaging protocol while also leveraging existing clinical records to enhance dataset diversity.

Data collection

To develop an AI-powered wound assessment tool, wound images and clinical data were gathered from multiple healthcare institutions using both retrospective and prospective methods. This dual approach ensured that the dataset incorporated both standardized imaging techniques and real-world clinical variability, resulting in a comprehensive and generalizable dataset. The final dataset included approximately 4,000 wound images, with 3,738 images collected retrospectively and 200 different wound images documented prospectively at multiple time points.

The study included only patients aged 18 years or older who demonstrated proficiency in the French language (as understanding French was required to comprehend the informed consent for prospective data collection) and who had an existing acute or chronic wound. Patients who did not meet the inclusion criteria or who lacked a valid informed consent were excluded from the study. The dataset included acute wounds, such as burns, surgical wounds, and traumatic wounds, as well as chronic wounds, including diabetic foot ulcers, pressure ulcers, arterial ulcers, venous ulcers, and mixed leg ulcers. This diversity was crucial to ensure that the AI model could generalize across different wound types. In addition, the images have been annotated for wound segmentation and tissue type classification.

Furthermore, the clinical data gathered were demographic information, wound characteristics, and treatment details. These data provided a holistic comprehension of the included wound types, which is crucial for developing a robust AI-powered wound assessment tool.

Prospective data collection

The prospective wound data collection was facilitated by imitoWoundR mobile application developed in collaboration with a Swiss startup (imito AG) specifically for this project. This application is designed to standardize imaging procedures across different clinical sites. Wound images have been captured during routine care at healthcare institutions in Geneva, Switzerland, with real-time feedback thanks to specific markers detection to ensure high-quality images. Before imaging, a calibration marker was placed near the wound to enable automated detection and measurement of 2D wound dimensions (i.e., width, height, and surface area). In addition, a ColorChecker® (Classic Mini) was systematically positioned adjacent to the wound in all prospective cases to support future color calibration studies. While both tools were consistently used during data acquisition to standardize image quality and enable subsequent analyses, color calibration was not applied in the current study’s AI model training or evaluation. A short 10- to 20-second video was first recorded, followed by a sequence of 20 high-resolution photographs from different angles. The application provided real-time quality feedback, ensuring that images have been only captured when the marker was correctly detected, reducing inconsistencies. At least one of these 20 images was selected, ensuring full visibility of the wound for complete and accurate segmentation of both wound and tissue types (see Fig. 1).

Fig. 1.

Tissue type annotation interface in the imitoWoundR App for prospective data collection. Wound border segmentation in green, granulation in red and slough in yellow

Images were taken using iPads (9th generation, iOS 16), while Structure Sensor (Mark II) technology enabled the acquisition of 3D wound scans. The entire procedure of video and image capture took approximately 2 to 5 minutes, depending on the size and location of the wound. Clinicians manually segmented wounds within the application using a touchscreen interface, outlining wound boundaries and identifying different tissue types. Segmentation could take between 5 to 15 minutes, depending on the wound size and the variety of tissue types present within the wound. In this study, the application was specifically developed for research purposes; therefore, it is important to note that segmentation will be performed automatically in the final product. This automation is expected to significantly reduce segmentation time, as only minimal manual adjustments may be required in the final application. In addition to imaging, demographic information, wound characteristics, and treatment details were recorded within the application. To protect patient confidentiality, all identifiable information was anonymized before being securely transmitted to a storage server managed by the HES-SO Geneva.

Retrospective data collection

To supplement the prospective dataset, wound images from existing clinical records were retrieved from multiple hospital databases. This retrospective method enabled the rapid accumulation of a large volume of wound images, covering a diverse range of wound types and patient demographics. Comprehensive patient demographic data were unavailable for certain cases, as half of the complete dataset was sourced from external collaborators without accompanying patient information.

A systematic search was conducted across institutional servers, specifically targeting departments involved in wound management. Only images that met resolution and inclusion criteria were selected, prioritizing those with multiple time points of the same wound to allow for tracking healing progression.

An anonymization process was applied to protect patient privacy. Any identifying features, including names, facial features, and tattoos, were censored before images have been uploaded to the secure HES-SO Geneva storage server. The identifying features could either be erased, blurred, or hidden using Adobe Photoshop. Once anonymized, wounds in the images have been manually segmented using VGG Image Annotator (VIA) and Adobe Photoshop, with trained annotators carefully outlining wound boundaries. Tissue types were classified into epithelialization, granulation, fibrin, slough, necrosis, and noble tissues such as bone, muscle, and tendon (see Fig. 2). A precise definition of these different tissue types is described in Table 1, which allowed us to have a standardization of the tissue identification for retrospective and prospective data collection:

Fig. 2.

Retrospective wound images and their respective wound tissue identification masks.. Left column: Three original images captured retrospectively by nurses on different health institutions. Right column: Corresponding tissue identification masks made with Adobe Photoshop (Epithelialization in pink, granulation in red, fibrin in yellow, slough in grey and necrosis in brown)

Table 1.

Tissue definition

Tissue	Definition
Epithelialization	This process refers to the restoration of exposed skin surface following an injury and is essential for wound closure. Epithelization can indicate an effective wound healing process. It is a complex biological mechanism involving specialized skin cells, which contribute to the formation of new epithelial tissue from the edge to the center of the wound [20]. Initially, this tissue color is deep pink, gradually becoming paler as it matures into new skin.
Granulation	When predominantly present, granulation tissue may indicate a favorable wound healing process. It is composed of key cells, which contribute to tissue and blood vessel regeneration. Revascularization allows oxygen to be transported within the wound, thereby promoting the healing process. Granulation tissue typically appears red to pink in color, depending on the capillaries network density within the wound [21].
Fibrin	This tissue type plays a crucial role in the wound healing process. In addition to facilitating hemostasis, it can also serve as a scaffold for cell migration, which is essential for the formation of new tissue and blood vessels [22]. Consequently, fibrin should not be removed during wound care. Its color ranges from yellowish to white. As fibrin thickens and becomes more viscous it transitions into slough, which impairs wound healing.
Slough	Slough formation results from an abnormal inflammatory response and can be an indicative feature for non-healing wounds. The presence of pathogenic bacteria and biofilm disrupts the wound healing process and can lead to cellular death. Slough should therefore be removed regularly to promote optimal wound healing. Its color can vary from pale yellow, to yellow-green, tan, or brown, depending on its composition. Proper identification of slough is critical for wound care, as it informs clinical decision-making regarding wound management [23]
Necrosis	The presence of necrosis indicates irreversible cellular death, often resulting from a prolonged lack of oxygen due to an interrupted blood supply within the wound [24]. Necrosis can be classified as dry or wet, depending on the wound type and location and can also be called eschar in literature when becoming dry and black. In contrast, the wet necrosis appears as a dark yellow to brown tissue and can resemble slough.
Noble tissues	Tissues, such as bone, muscle and tendons are not constitutional wound tissues, but they are typically protected by intact skin and defined as noble tissues. The presence of noble tissues can increase the risk of infection and necessitate meticulous wound care to preserve their integrity, while simultaneously removing devitalized tissue such as slough. The color of bones and tendons range from white to yellow, making differentiation from fibrin and slough complicated. In clinical practice, distinguishing between these tissues is facilitated by palpation, as bones and tendons are generally firmer than slough. On images, segmentation remains feasible since noble tissues should be found in specific anatomical locations and have distinctive textures.

Data labeling

Once anonymized, the ground truth annotations were defined as the manual segmentation masks created by a trained wound care researcher. The labels were then validated and corrected by four different experts in wound care with at least more than 10 years of clinical experience. For wound segmentation, these masks delineated the wound borders based on visual inspection of the images. For tissue type classification, expert annotators manually labeled regions corresponding to 7 predefined tissue classes based on standardized guidelines, namely: Epithelialization, Granulation, Slough, Fibrin, Necrosis, Bone and tendon, Other (for example: fat tissue, fascia, parts of dressing) (see Fig. 2). For the retrospective dataset, the annotations were performed using VGG Image Annotator (VIA) and Adobe Photoshop and were subsequently reviewed for consistency and accuracy by a second expert to ensure quality control. These expert-verified annotations served as the reference standard for model training and evaluation. A precise definition of these different tissue types is described in Table 1, which allows us to have a standardization of the tissue identification. Consistent tissue definitions were established at the start of the project and uniformly applied across both the retrospective and prospective datasets during the annotation process.

As for the prospective data, the annotation of the wound border and tissue types was performed with the imitoWoundR App developed specifically for this project to collect and label the data, as shown in Fig. 1.

Comparison of data collection methods

The combination of prospective and retrospective collection methods ensured that the dataset included both high-quality, standardized images and large-scale real-world data. Prospective collection provided consistent imaging conditions, while retrospective collection allowed for the rapid accumulation of diverse wound images from various clinical environments.

Notably, prospective images exhibited higher resolution, improved lighting conditions, and standardized color calibration, whereas retrospective images introduced greater variability in imaging conditions and occasionally lacked complete metadata. The trade-offs between these approaches underpin the importance of balancing standardization and scalability when developing AI-driven clinical tools.

Model Architecture, Training and Evaluation

We used the DeepLabV3 + architecture for both wound segmentation and tissue type segmentation tasks, leveraging a convolutional encoder-decoder structure with skip connections and a multi-scale Atrous Spatial Pyramid Pooling (ASPP) module, as shown in Fig. 3 [25]. For both tasks, the encoder was initialized with weights pre-trained on ImageNet to enable transfer learning.

Fig. 3.

Overview if the DeepLabV3 + architecture used for wound and tissue segmentation [25]

To improve generalization to diverse clinical imaging conditions, we applied the following data augmentation techniques during training using ImageDataGenerator:

• Random rotations (up to 120°)

• Horizontal and vertical shifts (up to 20%)

• Zooming (range 0.7–1.3 ×)

• Brightness variation (range 50%–150%)

• Constant fill mode (cval = 0) to handle padding during transformations

Wound segmentation

For the wound segmentation task, we employed a ResNet50 backbone. The model’s output is a single-channel segmentation mask, trained using the Sparse Categorical Crossentropy loss function and the Adam optimizer with a learning rate of 0.0001. The model was compiled with accuracy as the evaluation metric and trained for 150 epochs. A ModelCheckpoint callback was used to save model weights at each epoch.

Tissue type segmentation

For the tissue type multiclass segmentation task, the same DeepLabV3 + architecture was used but with a ResNet101 backbone to benefit from deeper feature extraction. The output layer used a softmax activation, with training based on the Sparse Categorical Crossentropy loss function and an Adam optimizer with a learning rate of 0.001. The model was trained for 150 epochs, using accuracy as the evaluation metric. The training callbacks included:

• ModelCheckpoint to save model weights at each epoch

• ReduceLROnPlateau to halve the learning rate when validation performance plateaued (patience = 5)

All training and evaluations were conducted using TensorFlow 2.14.

These segmentation models were trained using a stratified data partition of 70 % for training, 15% for validation and 15 % for testing.

Evaluation metrics

Both models were evaluated primarily using the following metrics to assess boundary quality and class overlap:

• Dice Coefficient (F1 Score):

  

  Measures the overlap between predicted and ground truth masks. A Dice score of 1 indicates perfect overlap.

• Intersection over Union (IoU):

  

  Also known as the Jaccard Index, it quantifies the ratio of intersection to union between prediction and ground truth.

• Recall (Sensitivity):

  

  Reflects the proportion of actual wound pixels correctly identified by the model.

Where:

• TP = True Positives (correctly predicted wound pixels)

• FP = False Positives (non-wound pixels wrongly predicted as wound)

• FN = False Negatives (wound pixels missed by the model)

These metrics were calculated on the validation set using the predicted binary masks compared to manual ground truth annotations.

Ethics approval and consent to participate

This study was conducted in accordance with the ethical standards of the Declaration of Helsinki and approved by the Research Ethics Committee of the Canton of Geneva, Switzerland (Reference number: 2022–00827). All participants received both written and verbal information about the study and provided written informed consent prior to participation. Participation was entirely voluntary, and individuals were informed of their right to withdraw from the study at any time without any consequences.

Results

The integration of retrospective and prospective data collection resulted in a high-quality, diverse wound image dataset, which is essential for improving AI performance under real-world conditions. A total of 4,000 wound images were collected, representing a wide variety of acute and chronic wounds, sizes, anatomical locations, and skin tones. All 4,000 images were annotated for wound segmentation, while a curated subset of 1,200 images was selected and annotated for tissue type classification. The reduction was due to the exclusion of cases such as stitches, healed wounds, skin tears, donor sites of skin grafts, wounds with skin grafts, deep tissue pressure injuries, and burns, which were not suitable for consistent tissue labeling. Prospective data contributed to greater standardization in imaging conditions, while retrospective data offered a broader range of real-world clinical variability. Prospective data set comprised 42 patients, with approximately 200 wound images collected at multiple time points. Tissue type classification utilized exclusively retrospective data, as modification of tissue segmentation in the prospective image set was not feasible within the constraints of imitoWoundR.

The combination of prospective and retrospective collection methods ensured that the dataset included both high-quality, standardized images and large-scale real-world data. Prospective collection provided consistent imaging conditions, while retrospective collection allowed for the rapid accumulation of diverse wound images from various clinical environments.

Notably, prospective images exhibited higher resolution, improved lighting conditions, and standardized color calibration, whereas retrospective images introduced greater variability in imaging conditions and occasionally lacked complete metadata. The trade-offs between these approaches underpin the importance of balancing standardization and scalability when developing AI-driven clinical tools. The final composition of the tissue type dataset is illustrated in Fig. 4, which shows the mean percentage distribution of tissue types across the annotated images. The most represented classes were slough (33.7%), granulation (27.0%), and epithelialization (16.5%), followed by necrosis (15.0%), fibrin (4.3%), other tissue (3.3%), and tendon/bone/muscle (0.2%). This class imbalance reflects clinical reality but presents challenges for training and evaluating AI models, particularly for rare tissue types.

Fig. 4.

Mean percentage distribution of annotated tissue types within the dataset. The figure illustrates the relative prevalence of each tissue category, highlighting variations in sample representation across the dataset. This distribution provides insights into the dataset's composition and potential class imbalance relevant for downstream analysis or model training

For the tissue segmentation task, we defined four main classes for model prediction: epithelialization, granulation, necrosis, and slough. This classification scheme was based on clinical relevance, data distribution, and the visual characteristics of wound tissues in real-world images.

These four classes were selected because they represent the most commonly observed and clinically actionable tissue types, directly informing wound care decisions such as the need for debridement, dressing selection, and healing stage assessment. In our annotation process, several less frequent or visually ambiguous categories, such as fibrin, tendon/bone/muscle, and other (fascia, parts of dressings, fat tissue), were grouped under the slough class. This decision was made for the following reasons:

• Visual similarity and annotation consistency: Certain tissue types (e.g., fibrin vs. slough) are difficult to distinguish in photographic images and often overlap in clinical descriptions, leading to inter-annotator variability.

• Class balance considerations: Rare classes like bone or donor site tissues were underrepresented in the dataset and contributed to significant class imbalance. Including them as separate labels would reduce model stability and performance.

This grouping strategy improves the model’s robustness and generalizability across varied wound types while maintaining clinical relevance in output interpretation.

The wound border segmentation model demonstrated robust performance, achieving an average DICE score of 92%, an Intersection-over-Union (IOU) score of 85%, and a recall rate of 90%. The comparison between the automated wound border segmentation and the ground truth is shown in Fig. 5. As a preliminary result for tissue segmentation, the model showed promising results, with a mean DICE score of 78% for tissue classification. However, performance varied across different tissue types, notably epithelialization and necrosis, highlighting the need for further refinement and targeted AI training. The individual DICE scores for the four predicted tissue classes were as follows: epithelialization (65.1%), granulation (81.3%), necrosis (71.8%), and slough (82.2%). These values illustrate that performance was highest for slough and granulation, which were also the most represented classes in the dataset. In contrast, epithelialization and necrosis, which are visually subtler and underrepresented, showed slightly lower segmentation performance.

Fig. 5.

Workflow of the automated wound border segmentation process compared with ground-truth annotations. From left to right, each set displays: the original wound image captured by nurses, the corresponding automated segmentation output, the binary wound mask, and the manually annotated ground-truth. This comparison visually assesses the segmentation model's performance against expert annotations

The successful integration of this AI model into a wound app has significantly advanced wound documentation by streamlining measurement processes, supporting treatment decisions, and enhancing patient outcomes. To enable efficient mobile deployment, the trained Keras model was converted to TensorFlow Lite (TFLite) format using the TFLiteConverter, with post-training quantization applied via tf.lite.Optimize.OPTIMIZE_FOR_SIZE. This optimization step quantized model weights to 8-bit integers, reducing the model size by approximately 75%, while preserving its architecture and predictive capacity.

This approach did not require retraining or calibration data and resulted in only a 0.3% decrease in Dice score, demonstrating excellent retention of performance. The quantized TFLite model for wound segmentation was deployed and released within the mobile app, where it achieved an average inference time of 0.3 seconds on standard smartphones. In contrast, the tissue type segmentation model has been deployed internally but not yet released, as further improvements are ongoing. These efforts focus on expanding the annotated dataset and refining label quality to enhance model robustness and clinical reliability. This balance of compact size, fast inference, and high accuracy underscores the model’s potential for seamless real-world integration and its suitability for point-of-care wound assessment.

Discussion

AI has shown potential in transforming wound care by automating assessment processes, improving accuracy, and supporting clinicians in diagnosis and treatment decisions. This study combined both retrospective and prospective data collection to develop a comprehensive dataset, ensuring standardization while capturing real-world variability. The combination of high-quality, controlled imaging and large-scale clinical data provided a robust foundation for AI-driven wound segmentation and tissue classification. In comparison with earlier work, the performance of our AI models is consistent with or superior to prior studies. Our wound segmentation model achieved a DICE score of 92% and IoU of 85%, while tissue classification reached a mean DICE score of 78%. These results are comparable to prior research using U-Net or similar architectures for wound segmentation on diabetic foot ulcers, which typically report DICE scores in the range of 81% to 88% [26, 27]. Similarly, Ramachandram et al. [28] demonstrated that mobile-deployable AI models can achieve wound segmentation performance close to 90%, validating our model’s feasibility for point-of-care use. For tissue classification, our model showed variable performance across tissue types, which aligns with findings from Howell et al. [7] and Chairat et al. [29] who also noted challenges in differentiating visually similar tissues. These comparative results reinforce the robustness and practical relevance of our approach while highlighting ongoing challenges in standardizing tissue classification. Accurate wound segmentation and tissue classification are critical for supporting clinical decision-making and evaluating treatment effectiveness. Precise delineation of the wound border enables consistent measurement of wound area, allowing clinicians to monitor healing progression over time and detect early signs of stagnation or deterioration. Similarly, automated tissue classification provides objective insights into wound bed composition, which are key indicators used in determining the appropriate course of treatment. For example, identifying slough or necrosis may prompt debridement, while detecting granulation or epithelialization can guide dressing selection and healing stage assessment. By generating reproducible and timely assessments, the model outputs aim to enhance clinical workflow and support standardized wound care aligned with evidence-based guidelines.

The dual data collection approach proved to be essential in overcoming limitations associated with previous AI-based wound studies. While some studies focused exclusively on diabetic foot ulcers, limiting generalizability, this dataset included a diverse range of wounds, enhancing its clinical applicability [26, 27, 30]. However, retrospective data collection introduced inconsistencies in image quality, lighting, and metadata completeness, similar to challenges reported by Ramachandram et al. [28]. These limitations were mitigated through preprocessing techniques, manual segmentation corrections, and real-time quality feedback mechanisms, which have been shown to enhance AI training efficiency [31]. Tissue segmentation posed challenges, particularly in differentiating fibrin from slough and necrosis. Similar difficulties have been reported in studies such as Chairat et al. and Howell et al., highlighting the subjectivity in clinician-defined tissue boundaries even with preliminary tissue definitions [7, 29]. Furthermore, tissue segmentation demonstrated consistency across multiple assessments of the same wound performed by the same expert, with the exception of epithelial tissue, which exhibited poor segmentation consistency [28]. Implementing advanced feature extraction techniques, including multi-modal imaging approaches like hyperspectral imaging or infrared thermography, could improve classification accuracy. The AI-powered wound assessment tool developed in this study has important implications for wound care, particularly in settings where specialist expertise is limited. Automating wound segmentation, and tissue classification can help non-specialist clinicians make more informed treatment decisions, reducing the need for frequent in-person assessments. In addition to this tool, we plan to develop a SG which could help improving non-specialist clinicians wound care knowledge with regular trainings. This aligns with the increasing need for AI-driven telemedicine solutions, especially in rural or resource-limited areas [32]. AI-based wound monitoring can also contribute to cost reduction by identifying high-risk wounds earlier, preventing complications, and optimizing treatment pathways.

Limitations and challenges

Despite the promising results, several challenges must be addressed to optimize AI-driven wound assessment models. One of the main limitations is the resource-intensive nature of prospective data collection, which required dedicated personnel, standardized imaging protocols, and real-time patient interactions. Compared to retrospective collection, which allowed for the rapid accumulation of large datasets, the prospective approach was time-consuming and required continuous clinical engagement. This limitation highlights the need for scalable AI training strategies that can efficiently leverage existing clinical datasets while still maintaining data quality for instance by employed semi-supervised learning approaches [33].

Another limitation relates to inter- and intrarater variability in wound and tissue segmentation. Although wound boundaries were consistently delineated using standardized annotation protocols, tissue type segmentation remains inherently subjective. Previous studies have suggested that ensemble learning techniques, where multiple AI models are trained on different segmentation strategies, can help reduce annotation bias [7, 28, 34]. Integrating such techniques into future iterations of AI training could improve segmentation consistency across different tissue types.

One of the key concerns for the widespread adoption of AI-driven wound assessment tools is regulatory approval and clinical validation. While the AI models developed in this study demonstrated high performance in wound segmentation and tissue classification, further clinical trials are needed to validate the model’s performance across different healthcare settings and patient populations. Given the increasing emphasis on explainable AI (XAI) in medical applications, future studies should also explore how AI-generated wound assessments can be integrated into existing electronic health record (EHR) systems in a way that is transparent, interpretable, and clinically actionable.

Future Research Directions

Further research should focus on enhancing AI model generalizability by incorporating additional wound types and imaging conditions. While this study included a wide range of acute and chronic wounds, certain wound types, such as radiation-induced wounds or rare dermatological conditions, were not well represented. Expanding the dataset to include these wound types would improve the AI model’s real-world applicability.

Improvements in tissue classification algorithms are also needed to enhance AI segmentation performance. One potential avenue is the use of multi-spectral imaging techniques, which have been shown to improve the differentiation of necrotic and granulating tissues by capturing wavelengths beyond visible light. Recent studies have demonstrated that machine learning models trained on multi-spectral imaging data outperform traditional RGB-based models in wound assessment [35, 36].

Another important direction for future work is the integration of AI-based wound assessment tools into mobile health (mHealth) applications [37]. Mobile applications are increasingly being used for remote patient monitoring, allowing patients to self-document wound images and share them with clinicians for remote evaluation. Implementing AI-driven wound analysis within an mHealth framework could significantly improve access to wound care in underserved communities, particularly in low-resource settings where specialist wound care expertise is scarce.

Conclusion

The study demonstrates that AI-powered wound assessment tools have the potential to revolutionize clinical wound care by enhancing accuracy, efficiency, and accessibility. The integration of prospective and retrospective data collection proved to be an effective strategy for building a high-quality, diverse dataset, enabling the development of robust AI models for wound segmentation, tissue classification. While challenges remain in standardizing tissue segmentation, improving AI model generalizability, and addressing clinical validation requirements, the results underscore the transformative potential of AI in wound management. Future research should focus on refining AI segmentation algorithms, expanding dataset diversity, developing a SG and integrating AI tools into mobile health platforms to further improve clinical decision-making and patient outcomes.

Author contributions

Alessio Stefanelli contributed to the conceptualization of the study, data collection protocol development, and drafted the initial version of the manuscript. Sofia Zahia was responsible for the methodological design, supervised the data analysis process, and critically revised the manuscript for intellectual content. Guillaume Chanel contributed to the development and optimization of the AI model and participated in the interpretation of the computational results. Rania Niri contributed to the development and optimization of the AI model and participated in the interpretation of the computational results. Swann Pichon contributed to the clinical design and supported manuscript editing. Sebastian Probst supervised the overall project, led the clinical aspects of wound care assessment, coordinated the collaboration between institutions, and reviewed and approved the final manuscript version. All authors read and approved the final manuscript.

Funding

This study was funded by InnoSuisse (59519.1 IP-LS).

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to restrictions related to patient confidentiality and institutional data protection policies.

Declarations

Ethics accordance statement

This study was conducted in accordance with the ethical standards of the Declaration of Helsinki and approved by the Research Ethics Committee of the Canton of Geneva, Switzerland (Reference number: 2022-00827). All participants received both written and verbal information about the study and provided written informed consent prior to participation. Participation was entirely voluntary, and individuals were informed of their right to withdraw from the study at any time without any consequences.

Consent for publication

Written informed consent for the use of anonymized data and images for scientific dissemination and publication purposes has been obtained from all participants included in the study.

Competing interests

The authors declare no competing interests.