Automated human cell classification in sparse datasets using few-shot learning.

Area of Science:

• Computational biology
• Machine learning

Background:

• Automating human cell classification is crucial for expediting analysis, often requiring trained professionals.
• Deep learning models show promise but demand extensive data, which is scarce for human cell datasets, leading to low performance.

Purpose of the Study:

• Investigate the feasibility of few-shot learning (FSL) techniques to reduce data requirements for accurate human cell classification.
• Evaluate current state-of-the-art FSL methods, explore backbone architecture and training scheme variations, and propose future research directions.

Main Methods:

• Evaluated state-of-the-art FSL techniques on human cell classification, training on non-medical data and testing on out-of-domain medical datasets.
• Experimented with EPNet, modifying backbone architectures, data augmentation, and training schemes to assess impact on performance.
• Analyzed performance drops and identified limitations of current FSL approaches for medical image classification.

Main Results:

• State-of-the-art FSL techniques experienced at least a 30% decrease in test accuracy when transitioning from non-medical to human cell datasets.
• EPNet and Reptile showed the best performance on the BCCD and HEp-2 datasets, respectively.
• Even with modifications, EPNet's accuracy dropped from 88.66% on non-medical data to a maximum of 44.13% on medical data.

Conclusions:

• Few-shot learning, in its current form, performs poorly on human cell classification tasks.
• Modifying existing network architectures did not effectively improve performance.
• Future research should prioritize enhancing out-of-domain robustness through optimization-based or self-supervised FSL techniques.