New publication at Nature Communications
Evaluating Normative Representation Learning in Generative AI for Robust Anomaly Detection in Brain Imaging" by Cosmin I. Bercea et al. has been accepted for publication in Nature Communications.
Our latest study explores the potential of normative representation learning in medical imaging, focusing on understanding typical anatomical patterns and detecting anomalies without expert labeling. The research introduces new metrics to evaluate normative learning in generative AI models, including advanced diffusion frameworks, and tests them across a wide range of brain pathologies.
A large multi-reader study compares these metrics to expert evaluations, demonstrating the models’ ability to detect diverse, unseen medical conditions effectively. The findings aim to improve the assessment of generative models and contribute to more robust, clinically relevant AI systems.
The code is available here: https://github.com/ci-ber/GenAI_UAD.
Check the pre-print for more information.
Cosmin I. Bercea
Research Scientist
I am a postdoctoral researcher specializing in vision and multimodal learning for medical image analysis, with the current focus on developing vision-language models for generative downstream tasks.