Voetencheck AI Triage
Master
Semester programme:Master of Applied IT
Client company:FHICT professorship, Fontys Paramedisch, lectoraat HIT, RondOm Lopen Groep
Luc Lehmkuhl
Project description
This project explores the applicabillity and impact of AI-assisted triage in diabetic foot care. The central research question asks whether modern AI architectures—specifically fine-tuned vision-language models (CLIP)—can outperform traditional CNNs and zero-shot approaches in classifying foot images into clinically meaningful severity levels. We built three pipelines (CNN, CLIP Zero-Shot, and Fine-Tuned CLIP), evaluated their performance using a 3-class severity scale (low, medium, high), and validated results against expert podiatrist annotations. The project aims to increase diagnostic efficiency and reduce workload in resource-constrained clinical settings.
Context
Diabetic foot complications are among the most severe and resource-intensive outcomes of diabetes, frequently leading to infections, hospitalizations, and even amputation. Early detection and triage are essential but typically depend on manual review of foot images, especially in telemedicine and under-resourced care settings. The work sits at the intersection of healthcare and AI, targeting clinical efficiency, consistency, and accessibility through automation. This project addresses a notable research gap: the lack of structured, AI-driven severity scoring systems tailored to podiatric imaging. By focusing on diabetic foot risk classification,I wanted to contribute to a domain with direct real-world applicability and measurable clinical value.
Results
The project produced three AI triage pipelines: a CNN baseline, a zero-shot CLIP model, and a fine-tuned CLIP model trained on expert-labeled diabetic foot images. Evaluation was conducted via 5-fold cross-validation on 80 foot images using macro-averaged metrics. The Fine-Tuned CLIP model outperformed all others, achieving an average accuracy of 82.86% (±6.97), F1-score of 0.54, and precision >80% across all severity levels. In contrast, CNN and Zero-Shot CLIP models underperformed, particularly in identifying high-risk cases.
Statistical analysis confirmed the significance of performance differences (p < 0.01), and heatmap visualizations demonstrated robust generalization across validation folds. From a TRL perspective, the solution reaches TRL 5—a validated prototype in a relevant clinical environment. The pipeline is reproducible, explainable, and readily deployable for further piloting.
This solution offers clear benefits: faster triage, reduced burden on podiatrists, and scalable deployment potential for remote or high-volume screening programs. The technical deliverables include trained models, clinical dashboards, evaluation reports, and a retrainable feedback loop—paving the way for future integration in diabetic foot care systems.
About the project group
I am a Software Engineer Bachelor from the Master of Applied IT program at Fontys University of Applied Sciences. With backgrounds in AI engineering and healthcare technology, I collaborated over a 20-week research period with healthcare professionals to develop and evaluate an AI-driven triage system for diabetic foot care. We adopted an agile, research-through-prototyping approach, working iteratively on model development, stakeholder engagement, clinical validation, and technical documentation. Throughout the project, we actively engaged with podiatric experts to ensure clinical relevance and maintained reproducibility standards through versioned code and dataset management.