Deep learning for diabetic foot ulcer: a systematic review of segmentation, classification and multimodal integration

Abstract Diabetic foot ulcers, DFUs, are a major complication of diabetes with substantial morbidity and amputation risk, motivating automated assessment systems that are accurate, reproducible, and deployable in point-of-care and remote-monitoring settings. This PRISMA 2020-compliant systematic review of 48 deep learning study configurations (2019–2024) critically examines deep learning approaches for DFU classification, lesion localization, segmentation, and multimodal fusion, with an explicit focus on engineering feasibility, evaluation comparability, and reproducibility. Following PRISMA 2020 [41], a structured search of major databases from 2019 to 2024 identified 48 eligible studies. In addition to synthesizing architecture families including convolutional neural networks, efficient compact networks such as EfficientNet and MobileNet variants, transformer and hybrid designs, and ensemble systems across sensing modalities including RGB, thermal or infrared, depth, multispectral or hyperspectral imagery, and clinical variables, we quantify methodological rigor using a 13-criterion checklist with a weighted methodological rigor score (WMRS) whose weighting scheme is justified in Section 2.4.1 and validated through a sensitivity analysis (Section 2.4.2; Table 2) showing study rankings are robust to weight choice (Spearman $$\rho \ge 0.92$$ across alternative schemes). The mean weighted score was 49.0% with a range of 23.8% to 95.2%. Reproducibility and clinical translation indicators were limited: only 27.1% (13 of 48) of studies reported external validation, 8.3% (4 of 48) released code, and 16.7% (8 of 48) described mobile or web deployment constraints, while 56.3% (27 of 48) used publicly available datasets. A descriptive cross-analysis (Section 6.7.1; Figure 5) shows that studies with external validation report on average $$\sim$$ 3 percentage points lower headline accuracy than those without, providing indirect evidence that internal-only evaluation may inflate reported performance. Reported performance varies with dataset composition and acquisition variability, validation design such as patient-level separation and cross-dataset testing, and metric selection spanning Dice, intersection over union, mean average precision, accuracy, and area under the curve. Multimodal systems combining RGB with thermal signals show promising gains for early physiological cues, but their reliability depends on calibration, drift control, modality registration, and standardized normalization suitable for resource-limited devices. We conclude with a deployment-oriented roadmap emphasizing leakage-resistant benchmarking, modality-specific acquisition and calibration guidance for thermal imaging, lightweight and energy-aware model design for edge devices, and clinically interpretable explainability to improve trust and adoption.