Deep learning-enhanced 3D real-time photoacoustic imaging using experimental ground truths obtained from fluctuation...

Area of Science:

• Biomedical optics
• Medical imaging
• Artificial intelligence in medicine

Background:

• 3D photoacoustic (PA) imaging faces visibility artifacts due to transducer limitations and sparse arrays.
• PA fluctuation imaging (PAFI) improves visibility but sacrifices temporal resolution.
• Deep learning (DL) shows promise for PA image enhancement but requires experimental training data.

Purpose of the Study:

• To develop a DL-based method for enhancing 3D PA images using experimental data.
• To train a 3D ResU-Net network using single-shot 3D PA images and PAFI images.
• To evaluate the performance of the DL-PAFI network for real-time, artifact-free 3D PA imaging.

Main Methods:

• Created an experimental training dataset from chicken embryo vasculature using single-shot 3D PA and PAFI images.
• Trained a 3D ResU-Net neural network on the experimental dataset.
• Evaluated the network's performance on new experimental test images and demonstrated real-time rendering capabilities.
• Tested the network's ability to predict *in vivo* images in mice.

Main Results:

• The DL-PAFI network effectively improved visibility and contrast in 3D PA images.
• Output image resolution was lower than PAFI, but training with experimental data alone yielded good performance.
• Pre-training with simulated data further improved overall accuracy.
• Demonstrated feasibility of real-time rendering and preliminary *in vivo* predictions in mice.

Conclusions:

• DL-based enhancement using experimental data can significantly improve 3D PA imaging quality.
• The trained network shows potential for real-time, artifact-free 3D PA imaging with sparse arrays.
• The method is adaptable for various *in vivo* applications, including cross-species predictions.