DOI: 10.1097/dm-2026-00020 ISSN: 2226-8561

Deep learning and computer vision for the assessment of granulation tissue in diabetic foot ulcers: A systematic review

Afif Singgih Prasetyo, Heri Kristianto, Retno Lestari, Rulli Rosandi, Andhika Yudistira
Show PDF Cite

Background:

Chronic hyperglycemia and oxidative stress lead to chronic, slow-healing wounds; this can result in diabetic foot ulcers (DFUs), which affect 6.3% of the diabetic population and lead to limb amputation in approximately 30% of cases. Because of their complex healing processes, including impaired angiogenesis and dysfunctional fibroblasts, a system of assessing granulation in DFUs is essential, especially in low-resource countries. Innovations in artificial intelligence have created models to automate the process of segregating wounds and to help quantify the amount of tissue. This review focuses on the assessment of DFU granulation using deep learning models and the barriers to and challenges in clinical application and implementation.

Methods:

This systematic review followed PRISMA 2020 guidelines and was registered in Open Science Framework Registries (https://osf.io/rkpv8). PubMed, ScienceDirect, Scopus, and IEEE Xplore were searched for studies published between January 2015 and March 2025 evaluating deep learning or computer vision models for DFU granulation tissue segmentation or classification. Three reviewers independently performed screening, data extraction, and quality assessment using an adapted QUADAS-2 framework covering dataset selection, index test, reference standard, and validation strategy. Out of 766 identified records, 16 studies meeting the quantitative eligibility criteria were included. Due to substantial methodological heterogeneity, a narrative synthesis was performed instead of a meta-analysis.

Results:

Sixteen studies were included. All studies reported using at least one public or private dataset and at least one Diabetic Foot Ulcers Grand Challenge (DFUC) 2022 dataset. The most common frameworks were U-Net and its derivatives ( n = 7), followed by standard convolutional neural networks (CNNs), EfficientNet models, and hybrid CNN-transformer models. The researchers reported Dice similarity coefficient (DSC) values between 0.88 and 0.94 and IoU values between 0.82 and 0.91. All area under the curve (AUC) values, which range between 0 and 1 and where values closer to 1 indicate better performance, were greater than 0.85, thus indicating good performance. The highest segmentation performance was achieved with optimized EfficientNet-B3, DeepLab V3+, DoubleU-Net, and transformer models. Image quality, expert annotations, preprocessing, and validation approached performance in a systematic manner. Most of the studies reported a low bias risk; however, there were concerns about the study datasets and the external validation.

Conclusion:

Models based on U-Nets, when coupled with ResNets, EfficientNets, and transformers, demonstrate high precision in diagnosing the granulation level of DFUs and can have significant positive effects on digital wound care and telemedicine. However, their lack of multicenter validation and retrospective, single-center designs currently limits the clinical application. To be able to use these models in daily clinical DFU management, large prospective multicenter studies must be conducted.

More from our Archive