Improving the quality of patient-captured diabetic foot images using AI-based guidance and validation

Project description

This project's goal is to support both podiatrists and patients in remote diabetic-foot monitoring. Patients take photos of their own feet at home so clinicians can screen for early signs of diabetic foot disease or other complications without frequent clinic visits.

However, many of these images are too blurry, dark, or poorly framed to be clinically useful. This project evaluates whether an AI-based “quality gate” can assess image quality at capture time and guide patients to retake better photos. The system combines an IQA model with simple OpenCV checks, allowing only images that meet podiatrist-defined criteria to proceed.

The key challenge is keeping this gate strict enough to prevent poor-quality images while remaining easy for patients to use.

Context

This project is set in the healthcare and digital health domain, specifically within remote diabetic foot care and telemedicine.

Healthcare providers increasingly rely on patient-captured smartphone images to monitor foot health outside the clinic, aiming to improve access to care and reduce in-person visits.

This shift introduces new challenges around data quality, clinical reliability, and efficient use of clinician time in remote care workflows.

Results

The project resulted in a functional mobile prototype that embeds two AI models within a remote diabetic foot monitoring workflow. One model assesses image quality at capture time, and only images that pass this check are forwarded to a second model for preliminary foot condition classification. The image quality model achieved a validation accuracy of 0.777 and largely aligned with podiatrist judgement, with disagreements mainly limited to borderline cases.

Applying a conservative quality gate helped prevent poor-quality images from being analysed further, supporting clinical safety. Clinicians reported that accepted images were more consistent and required fewer manual rejections, and that the in-app guidance reduced both retakes and the time needed to capture a usable image.

Overall, the findings indicate technical feasibility and practical relevance at roughly TRL 5–6, while highlighting the need for further patient-based evaluation and robustness testing before real-world deployment.

About the project group

My name is Shanessa Kostaman, and I am a Master’s student with a Bachelor’s in Software Engineering from Fontys ICT. I worked on this project full-time (40 hours per week) over a six-month period.

My process was split between technical development at home and regular meetings at school to discuss requirements and progress with stakeholders.

Most importantly, I was able to spend time at a podiatry clinic to test the application with real patients, which allowed me to refine the project based on actual user feedback.