Automated Classification of Apoptosis in Phase Contrast Microscopy Using Capsule Network

Area of Science:

• Cell Biology
• Biomedical Imaging
• Machine Learning

Background:

• Accurate apoptosis classification is crucial for cell biology research.
• Current deep convolutional neural networks (CNNs) struggle with part-whole relationships, necessitating large datasets.
• This limitation hinders robust generalization in apoptosis classification.

Purpose of the Study:

• To propose an efficient variant of capsule networks (CapsNets) for apoptosis classification.
• To leverage temporal information in microscopy videos using a recurrent CapsNet.
• To demonstrate the superiority of CapsNets over CNNs, particularly with limited training data.

Main Methods:

• Development of an efficient capsule network (CapsNet) variant.
• Integration of CapsNet with a bi-directional long short-term recurrent structure to form a recurrent CapsNet.
• Experimental evaluation on target cell apoptosis classification using microscopy videos.

Main Results:

• Proposed CapsNets achieve competitive performance in apoptosis classification.
• CapsNets significantly outperform CNNs when training sample size is small.
• Recurrent CapsNet achieves 93.8% accuracy and more consistent predictions on microscopy videos, utilizing temporal information.

Conclusions:

• Capsule networks provide an efficient and effective alternative to CNNs for apoptosis classification.
• Recurrent CapsNets enhance classification accuracy and temporal consistency by incorporating video data.
• This approach holds promise for advancing cell biology research through improved automated analysis.