2024/01/11 | Research | Artificial Intelligence
In medical image data sets with limited labeled data, active learning approaches can help reduce annotation costs and boost performance of computer-aided diagnosis systems. Based on its previous findings employing interpretability information as inductive bias and as criteria to select informative samples in active learning, the Medical Image Analysis lab has now developed a new graph-based transformer and data augmentation active learning framework for multi-label xray classification.
In collaboration with Monash University, Australia, and EPFL, the group proposed to use graph attention transformers (GATs) that enable learning more accurate data representations. A novel "multi-label informativeness score" derived from GATs quantifies the importance of each sample based on multi-label interactions. This was supplemented by a new data augmentation approach to generate novel transformations such that new synthetic images are ensured to also be informative compared to their base image, while enforcing class label preservation and redundancy avoidance.
When benchmarked nine other SOTA methods on two publicly available chest X-ray datasets and four other multi-class MedMNIST datasets, the proposed framework "GANDALF" (Graph-based transformer and Data Augmentation Active Learning Framework), outperformed the other methods for multi-label classification. This is due to GANDALF´s (a) use of a better method to select most informative samples based on multi-label interactions, and (2) its leveraging data augmentation to generate more informative synthetic samples based on multi-label interactions.
Link to the study
Medical Image Analysis