Model-based deep learning framework for accelerated optical projection tomography.

Area of Science:

• Biomedical imaging
• Computational imaging
• Deep learning applications

Background:

• Optical projection tomography (OPT) is crucial for 3D biological imaging.
• Traditional OPT reconstruction is time-consuming.
• Reducing acquisition and reconstruction time is essential for high-throughput biological studies.

Purpose of the Study:

• To introduce a novel model-based deep learning algorithm for accelerated optical projection tomography (ToMoDL).
• To reduce acquisition and reconstruction times in OPT.
• To compare ToMoDL performance against established reconstruction methods.

Main Methods:

• Developed a model-based deep learning algorithm (ToMoDL) incorporating data consistency and artifact removal via convolutional neural networks.
• Included a preprocessing step for accurate sample alignment.
• Trained the algorithm on zebrafish developmental stages to minimize mean square error.
• Validated using cross-validation against filtered backprojection, compressed sensing, and a direct deep learning method.

Main Results:

• ToMoDL achieves reconstruction quality comparable to or better than existing methods.
• For highly undersampled datasets, ToMoDL and a U-Net method yield comparable images.
• ToMoDL demonstrates superior performance with limited training data due to fewer trainable parameters.

Conclusions:

• ToMoDL offers a significant acceleration of optical projection tomography reconstruction.
• The method is robust and performs well across various undersampling scenarios.
• ToMoDL is particularly advantageous in situations with limited training datasets, making it suitable for resource-constrained research.