• High AI Prediction Accuracy: ADimaster achieves an area under the curve (AUC) score of 87.9% in estimating the severity level of AD and provides highly accurate assessments across diverse ethnic groups, ensuring inclusivity and fairness. Additionally, it performs consistently with images from  a wide range of smartphone cameras with different specifications, making it a reliable solution for patients.
• Near Real-time Inference: With a rapid response time of under 0.2 seconds, patients can receive real-time, AI-driven severity assessments that facilitate prompt decision-making and support consistent self-monitoring, empowering them to manage their condition without delay.
• Comprehensive Disease Progress Tracking: ADimaster provides valuable insights into the severity of conditions, allowing patients to monitor their progress over time. This feature could also help patients and healthcare providers identify the most effective medications or therapies tailored to specific conditions, leading to better disease management outcomes.

The Science Behind

ADimaster leverages two key technologies: (1) feature disentanglement approach and (2) sharpness-aware optimiser. The feature disentanglement approach effectively separates domain-invariant features from domain-specific ones, enabling AI to generalise more efficiently across diverse input conditions. 

Fig 1. An overview of the feature disentanglement-based solution for domain generalisation

Fig 2. Comparison between stochastic gradient descent optimisation and our sharpness-aware optimisation method

To ensure consistent performance, the team developed diverse datasets that comprise images of all skin tone from multiple ethnic groups, captured under different lighting conditions using smartphone cameras with varying specifications. By leveraging robust machine learning techniques, the deep learning model was trained on this rich dataset, achieving high accuracy and robustness. Validation through experimental studies and on-site trials has proven that ADimaster demonstrates reliably across a broad range of self-monitoring scenarios.