A comparative study of cell classifiers for image-based high-throughput screening.

Area of Science:

• Computational Biology
• High-Throughput Screening (HTS)
• Machine Learning in Cell Biology

Background:

• High-throughput screening (HTS) generates millions of cell images requiring phenotype classification.
• Manual cell classification is infeasible due to the large data volume.
• Iterative training data improvement is crucial for accurate classification models.

Purpose of the Study:

• To compare the computational performance of various classification methods for cell phenotype identification.
• To evaluate gentle boosting, CellProfiler Analyst (CPA), Support Vector Machines (SVM), and Linear Discriminant Analysis (LDA).

Main Methods:

• Comparative computational performance analysis of classification algorithms.
• Testing on two distinct cancer cell datasets: HT29 and HeLa.
• Evaluation metrics include classification accuracy and processing time.

Main Results:

• For HT29 cells, linear SVM, radial basis function (RBF) SVM, and gentle boosting showed similar performance, with linear SVM being the fastest.
• The performance difference between linear SVM and RBF SVM on HT29 data was minimal (0.42%).
• On HeLa cells, RBF SVM achieved superior performance, outperforming linear SVM by an average of 1.41%.

Conclusions:

• Linear SVM is recommended for iterative training data refinement in HTS.
• RBF SVM is proposed as the optimal classifier for the final classification of unlabeled cells in large HTS datasets.