Sphere estimation network: three-dimensional nuclei detection of fluorescence microscopy images

Area of Science:

• Biomedical Imaging
• Computational Biology
• Deep Learning

Background:

• Fluorescence microscopy is vital for visualizing 3D subcellular structures.
• Two-photon microscopy enables deeper tissue penetration for imaging.
• Accurate nuclei detection is crucial for tissue analysis but challenging due to spatial variability.

Purpose of the Study:

• To develop a 3D convolutional neural network (SphEsNet) for simultaneous nuclei localization and size estimation.
• To eliminate the need for additional post-processing steps in nuclei analysis.
• To improve the efficiency and accuracy of extracting nuclear characteristics from microscopy data.

Main Methods:

• A 3D convolutional neural network, SphEsNet, with two branches for center coordinate localization and radius estimation was developed.
• Synthetic microscopy volumes generated via a cycle-consistent adversarial network were used for training.
• Three SphEsNet models were trained and tested on fluorescence microscopy datasets from rat kidney and mouse intestine.

Main Results:

• SphEsNet successfully detected nuclei across various locations and sizes in real microscopy data.
• The method outperformed existing techniques in object-level precision, recall, and F1 score.
• The model achieved an F1 score of 89.90%.

Conclusions:

• SphEsNet enables simultaneous nuclei localization and size estimation without post-processing.
• The network has the potential to extract more comprehensive information from nuclei in fluorescence microscopy images.
• This advancement can aid in clinical and research tissue analysis.