Professor Oliver Diaz attended the 17th International Workshop on Breast Imaging (IWBI2024) at the University of Chicago, where he presented a study entitled “Mitigating annotation shift in cancer classification using single-image generative models”. The research was conducted in collaboration with Marta Buetas Arcas, Dr Richard Osuala, Prof Karim Lekadir and the Artificial Intelligence Medicine Lab. The poster attracted a lot of attention and stimulated insightful discussions on the topic. The committee recognised the quality of the study by awarding it the First Runner-Up Poster Award at IWBI2024.
About the study
AI is an important tool for radiologists in the detection and diagnosis of breast cancer. However, can image features be used to predict cancer risk before biopsy?
The main challenge of AI in this area is that it is often limited by limited and costly data annotation shifts. This study simulates, analyses and mitigates the impact of these shifts on breast cancer classification.
Firstly, a highly accurate model has been developed that can predict cancer risk and effectively discriminate between benign and malignant lesions. Second, the performance of the model was used to quantify the impact of the annotation shift. The study revealed a significant impact of annotation shift on multiclass classification performance, especially for malignant lesions. Thus, a training data augmentation approach based on single image generative models for the affected class was proposed, which requires only four in-domain annotations to significantly mitigate annotation shift. Finally, performance was further improved by proposing and validating an ensemble architecture based on multiple models trained under different data augmentation regimes.
The paper has been published by SPIE as part of the conference proceedings of IWBI2024.
An open access pre-print version of the paper is available via arXiv.
The study was developed with support of EU Horizon 2020 project EuCanImage, EU Horizon Europe project RadioVal, and Spanish funded project FUTURE-ES.
