Deep learning optimization for small object classification in lensfree holographic microscopy

Area of Science:

• Optics and Photonics
• Biomedical Engineering
• Machine Learning

Background:

• Lensfree holographic microscopy offers a cost-effective, high-resolution imaging solution for large fields of view.
• This technique, when combined with automated image processing and neural networks, enables efficient biomolecular sensing using labeled micro- and nano-beads.
• Existing neural network applications in microscopy often focus on image reconstruction, leaving the optimal architecture for small object classification in holographic sensors undetermined.

Purpose of the Study:

• To investigate the performance of shallow convolutional neural networks (CNNs) for small object classification in lensfree holographic microscopy-based sensors.
• To thoroughly analyze the impact of various layers (dropout, convolutional, normalization, pooling, activation) and hyperparameters (dropout fraction, filter number/size, stride, padding) on CNN performance.
• To identify the most critical architectural choices for maximizing classification accuracy in this specific application.

Main Methods:

• A shallow convolutional neural network (CNN) architecture was designed and applied to classify small objects in lensfree holographic microscopy images.
• Systematic evaluation of different CNN layers, including dropout, convolutional, normalization, pooling, and activation functions.
• Investigation of key hyperparameters such as dropout fraction, filter number and size, stride, and padding to understand their influence on accuracy.

Main Results:

• The developed CNN achieved an overall classification accuracy of approximately 83% for small objects in holographic images.
• Analysis revealed that the choice of activation layer had the most significant impact on maximizing the network's accuracy.
• Specific layers and hyperparameters were found to influence network performance, providing insights into optimal CNN design for this task.

Conclusions:

• Shallow convolutional neural networks are effective for small object classification in lensfree holographic microscopy sensors.
• The selection of the activation layer is a crucial factor for optimizing accuracy in these networks.
• The findings offer valuable guidance for researchers developing neural networks for similar holographic-based sensing and classification tasks.