Small training dataset convolutional neural networks for application-specific super-resolution microscopy.

Area of Science:

• Microscopy
• Machine Learning
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) enhance microscopy, but require large datasets for training.
• Acquiring extensive datasets is a significant bottleneck in developing effective ML models for microscopy.

Purpose of the Study:

• To demonstrate a novel neural network modification enabling super-resolution (SR) image estimation from diffraction-limited (DL) microscopy images using small datasets.
• To overcome the data acquisition bottleneck in training ML models for microscopy.

Main Methods:

• Incorporation of "dense encoder-decoder" (DenseED) blocks into existing SR ML network architectures.
• DenseED blocks utilize dense layers to concatenate features, improving information flow.
• Training fully convolutional networks (FCNs) with DenseED blocks on a small dataset (15 FOVs) of human cells and endothelial cells.

Main Results:

• ML models with DenseED blocks accurately estimate SR images from DL images when trained on small datasets.
• Models without DenseED blocks fail to produce accurate SR images under similar training conditions.
• Networks incorporating DenseED blocks demonstrated significant improvements in peak signal-to-noise ratio (PSNR) and resolution.

Conclusions:

• DenseED blocks enable accurate SR image extraction even with minimal training data (15 FOVs).
• This approach facilitates the use of smaller, application-specific datasets for microscopy ML.
• The method shows potential for application in other imaging modalities like MRI and X-ray.