Skin Cancer Detection With AI

Project description

AI Feedback Loop

What if dermatologists could directly improve the accuracy of 'AI skin cancer detecting models' by integrating their expertise during their clinical workflows? This project explores the design of an AI feedback loop that allows continuous learning through dermatologist annotations and corrections. Unlike traditional AI systems that rely on outsourced annotations, this feedback loop captures clinically relevant data, addressing concerns about dataset diversity and expertise. By allowing dermatologists to refine AI predictions, the system creates trust and ownership while laying the foundation for more reliable diagnostic tools.

System Architecture

To support this vision, a modular system architecture was proposed, ensuring secure data handling, compliance with the EU AI Act and GDPR, and adaptability for diverse clinical settings. The architecture separates preprocessing, analysis, feedback collection, and model retraining components, making it scalable and flexible. While still in exploratory stages, this research highlights a pathway for integrating AI into dermatology, ensuring it aligns with the real-world needs of clinicians and patients.

Context

AI Feedback Loop

Current AI systems face limitations, such as reliance on outsourced annotations and a lack of diverse datasets. These gaps reduce the reliability and applicability of AI in real-world clinical settings. This research investigated whether an AI feedback loop could integrate dermatologist expertise into AI workflows, allowing continuous learning and improvement. Dermatologists provided annotations like bounding boxes and corrections, offering valuable data to refine the AI model. The system was designed to complement existing workflows, ensuring it could improve efficiency without disrupting patient care.

System Architecture

A modular architecture was developed to support the feedback loop, focusing on secure data processing and compliance with GDPR and the EU AI Act. The architecture's design ensures adaptability for various clinical settings, from single-practitioner clinics to larger healthcare networks. While the research remains exploratory, it provides a realistic framework for how AI systems can evolve through expert input, addressing key challenges in accuracy and applicability. This work not only meets the needs of MohsA Huidcentrum but also positions itself as a possible solution for other clinics.

Results

AI Feedback Loop

Can a collaborative AI system revolutionize skin cancer detection? This research suggests it potentially can. The system’s ability to gather feedback was tested in a clinical context. While the research did not measure long-term impacts on model accuracy, the prototype demonstrated feasibility and practical potential, gaining valuable insights for future implementations and refinements.

Dermatologists annotated images generating clinically relevant feedback that avoided reliance on outsourced annotations. The system achieved an average System Usability Scale (SUS) score of 85%, indicating acceptance among participants. Task efficiency was measured at 0.33 tasks per second, reflecting the system's ability to handle feedback and annotations efficiently. Simpler tasks, such as confirming predictions, were completed faster than more complex ones like creating bounding box annotations.

System Architecture

The prototype successfully demonstrated its ability to process and analyze images with accuracy and efficiency. Feedback from dermatologists was seamlessly incorporated into the retraining process. To support the clinic in adopting the system, detailed documentation was provided, including architectural designs, compliance strategies, and technical specifications.

Key results include reducing the time required for skin cancer diagnosis, improving AI accuracy over time, and securely managing clinical data. The modular architecture separates key components like preprocessing, feedback collection, and retraining, making the system adaptable for larger dermatology settings.

Impact and Future Potential

This project addresses current research gaps in AI for dermatology, such as lack of dataset diversity, clinical relevance, and compliance with legal frameworks. While exploratory, the research provides early valuable insights for developing scalable and responsible AI systems. By integrating human expertise into AI workflows, this work sets the stage for more accurate, reliable tools in healthcare. The AI feedback loop has the potential to reshape how dermatology—and other medical fields—approach AI based diagnosis and patient care.

About the project group

The group's recent educational background

Jen Nicolaes - Software Developer

Education: Fontys ICT & Software Graduate | Followed Advanced Software course
Specialization: Game Development
Minor: Global Acting IT
Currently Exploring: AI Software Architecture in Healthcare

Knarik Poghosyan - Software Developer & Media Designer

Education: Fontys ICT Graduate
Specialization: Cyber Security
Additional Programs: Tech Entrepreneurship, Cognitive Behavioral Research, Delta
Currently Exploring: Human-AI Communication in Healthcare