Unsupervised feature extraction from a foundation model zoo for cell similarity search in oncological microscopy across devices

Acquiring high-quality datasets in medical and biological research is costly and labor-intensive. Traditional supervised learning requires extensive labeled data and faces challenges due to diverse imaging equipment and protocols. We propose Entropy-guided Weighted Combinational FAISS (EWC-FAISS), using foundation models trained on natural images without fine-tuning, as feature extractors in an efficient and adaptive k-nearest neighbor search. Our approach shows superior generalization across diverse conditions, achieving competitive performance compared to fine-tuned DINO-based models and NMTune, whilst reducing computational demands. Experiments validate the effectiveness of EWC-FAISS for efficient and robust cell image analysis.