Support on data selection and AI model development for longitudinal domain shift detection for deep learning in digital pathology

Ongoing:	Yes
Start:	2023-12-01

Ensuring consistent and reliable performance of AI models in healthcare is critical, particularly given the challenges posed by domain shifts, where differences in data distribution between training and testing datasets can lead to a decline in model effectiveness. While domain shifts have been extensively studied in other areas, such as radiology, there is limited research on how this issue affects AI models in digital pathology.

The project aims to explore these longitudinal domain shifts through the lens of tumor classification in histopathology, using a large skin dataset to train an AI model for diagnosing benign lesions, malignant melanoma, basal cell carcinoma, and squamous cell carcinoma. The project will develop an unsupervised method to detect performance shifts, which could automatically alert quality assurance experts or pathologists when model accuracy drops beyond acceptable limits. This proactive approach will help maintain model efficacy in clinical settings, ensuring patient safety over time.

AIDA Data Hub supports the projects in data selection, tumor classifier training pipeline and advice on domain shift detection.