Single-shot self-supervised object detection in microscopy.

Area of Science:

• Microscopy
• Machine Learning
• Image Analysis

Background:

• Object detection in digital microscopy is crucial.
• Classical methods are limited.
• Machine learning requires extensive labeled or simulated data, which is often difficult to obtain.

Purpose of the Study:

• To develop a deep learning method for accurate object detection in microscopy.
• To overcome the reliance on labeled or simulated data.
• To detect microscopic objects from single, unlabeled images.

Main Methods:

• Proposed LodeSTAR (Localization and detection from Symmetries, Translations And Rotations), a deep learning approach.
• Exploited inherent roto-translational symmetries of microscopic objects.
• Applied the method to analyze challenging experimental data.

Main Results:

• LodeSTAR achieved sub-pixel accuracy in object detection.
• Outperformed traditional methods on accuracy, even with dense cells or noisy backgrounds.
• Demonstrated capability to measure additional properties like vertical position and polarizability using symmetries.

Conclusions:

• LodeSTAR offers a novel, data-efficient solution for object detection in digital microscopy.
• The method's reliance on symmetries enables accurate analysis of unlabeled, experimental images.
• Potential for broader applications in quantitative microscopy and material science.