Efficient Model Training and Few-Shot Learning in Biomedical Imaging

The field of biomedical imaging is moving towards efficient model training and few-shot learning, with a focus on developing innovative methods to reduce computational time and resources. Recent studies have introduced coreset selection methods, which aim to select a representative subset of the dataset for training and hyperparameter search. These methods have shown promising results in improving model performance while reducing computational overhead. Additionally, few-shot learning approaches have been explored, enabling models to recognize new classes from only a small number of labeled examples. Noteworthy papers in this area include HyperCore, which proposes a robust and adaptive coreset selection framework, and MetaChest, which presents a large-scale dataset for generalized few-shot learning of pathologies from chest X-rays. Foundation models have also been shown to be effective in few-shot anomaly detection, with FoundAD achieving competitive performance using substantially fewer parameters than prior methods.

Sources

Coreset selection based on Intra-class diversity

HyperCore: Coreset Selection under Noise via Hypersphere Models

SubZeroCore: A Submodular Approach with Zero Training for Coreset Selection

Performance-Efficiency Trade-off for Fashion Image Retrieval

Towards generalizable deep ptychography neural networks

MetaChest: Generalized few-shot learning of patologies from chest X-rays

David and Goliath in Medical Vision: Convolutional Networks vs Biomedical Vision Language Models

Assessing Foundation Models for Mold Colony Detection with Limited Training Data

Foundation Visual Encoders Are Secretly Few-Shot Anomaly Detectors

Built with on top of