Principles of PSSR

Note

This section can be considered an abbreviation of important concepts covered in the full PSSR paper, Deep learning-based point-scanning super-resolution imaging.

This page will soon be updated to contain information contained within the PSSR2 preprint, PSSR2: a user-friendly Python package for democratizing deep learning-based point-scanning super-resolution microscopy.

Why Use PSSR

Data aquisition is often costly and time consuming in microscopy, requiring the balance of numerous parameters to capture an ideal image. Additionally, high resolution/exposure data can be destructive of samples, especially when measuring fluorescence.

PSSR allows these limitations to be mitigated by super-resolving a low-resolution image to a high-resolution image.

Consider a set of high-low-resolution image pairs where the high-resolution images have 4 times resolution of their corresponding low-resolution images. Although the high-resolution images may contain 16 times the amount of pixels, they do not contain 16 times the amount of information about the sample. For example, some pixels may be redundant in the high-resolution images while their structure may be easily inferable in the low-resolution images.

This allows super-resolution techniques to super-resolve low-resolution images into their high-resolution counterparts without hallucinations (when benchmarked against human experts) provided a suitable model is trained.

The Crappifier

PSSR is based off conventional super-resolution techniques, but differs from the norm in its implementation. Perhaps the biggest principle of PSSR is that of the Crappifier.

In order to train a super-resolution model, high-low-resolution image pairs must be provided to the model as the input and target respectively. However, due to the difficulty and costs of acquiring high-low-resolution image pairs for training, this approach is not just impractical, but may be impossible for non-static live samples such as subcellular organelles.

To circumvent this, a Crappifier is used during training to computationally degrade high-resolution images into their low-resolution image counterparts. This process of crappification does not require low-resolution images to be acquired for training.

It first downsamples the high-resolution image to a lower resolution, before injecting a layer of noise that approximates the sampling noise of a true low-resolution image.

Although the model is not trained on true low-resolution images, the generated high-low-resolution image pairs are representative enough of the ground truth that the model can make accurate inferences on real world data.