Cell Image Classification: A Comparative Overview.

Area of Science:

• Cell biology
• Medical imaging
• Computational pathology

Background:

• Cell image classification is vital for drug discovery, protein localization studies, and cancer diagnosis.
• Existing methods include numerical feature extraction, neural networks (NNs), and transport-based morphometry (TBM).

Purpose of the Study:

• To compare the performance of numerical feature extraction, NNs, and TBM for cell image classification.
• To identify the strengths and weaknesses of each method across different datasets.

Main Methods:

• Comparative analysis of three cell image classification approaches: numerical feature extraction, NNs, and TBM.
• Evaluation on four publicly available cell imaging datasets.

Main Results:

• Numerical features demonstrated superior discriminative power for most classification tasks.
• NNs achieved state-of-the-art performance on datasets with abundant training samples.
• Limited training data did not consistently benefit from NN data augmentation or newer architectures.
• TBM methods offer interpretability through invertible classification functions.

Conclusions:

• The choice of cell image classification method depends on dataset size and the need for interpretability.
• Numerical features are robust for general classification, while NNs are powerful for large datasets.
• TBM provides valuable insights for understanding classification decisions.