Classifying changes in LN-18 glial cell morphology: a supervised machine learning approach to analyzing cell...

Area of Science:

• Cell biology
• Bioinformatics
• Machine learning

Background:

• Cell-based research generates large image datasets for treatment evaluation.
• Traditional end-point assays are error-prone, and existing computational methods are complex.
• Need for simple, powerful machine learning frameworks for cell microscopy data analysis.

Purpose of the Study:

• To evaluate existing machine learning frameworks for cell microscopy image analysis.
• To detail a machine learning pipeline for pixel-based and object-based classification.
• To compare the performance of three classifiers: random forest (RF), sequential minimal optimization (SMO), and Bayesian network (BN).

Main Methods:

• Image preprocessing using smoothing and contrast enhancement in FIJI.
• Pixel-based classification using the Trainable Waikato Segmentation (TWS) tool.
• Object-based classification using the Waikato Environment for Knowledge Analysis (WEKA) interface.
• Performance evaluation of RF, SMO, and BN classifiers using WEKA's experimental explorer.

Main Results:

• Bayesian network (BN) showed the lowest classification accuracy for both pixel-based and object-based models.
• Object-based SMO classifier demonstrated the best performance with a mean absolute error of 0.05.
• TWS and WEKA facilitate classifier creation and training but are not ideal for large datasets.

Conclusions:

• Machine learning frameworks offer a viable alternative to traditional methods for cell microscopy analysis.
• Object-based SMO classification provides a high-performing and accessible approach.
• While TWS and WEKA are user-friendly, scalability for large image datasets requires further consideration.