Deep image prior for undersampling high-speed photoacoustic microscopy.

Area of Science:

• Biomedical Imaging
• Optics and Photonics
• Machine Learning Applications

Background:

• Photoacoustic microscopy (PAM) combines light and ultrasound for imaging.
• High-speed PAM often undersamples data to increase imaging speed.
• Existing deep learning (DL) methods for sparse PAM require extensive training data and time.

Purpose of the Study:

• To introduce deep image prior (DIP) for improving image quality in undersampled PAM.
• To demonstrate a DL approach that bypasses the need for pre-training and ground truth data.
• To enable fast and flexible image enhancement for high-speed imaging.

Main Methods:

• Implementation of deep image prior (DIP) for image reconstruction.
• Application of DIP to sparsely sampled, high-speed photoacoustic microscopy data.
• Evaluation of DIP performance against interpolation and supervised DL methods.

Main Results:

• Substantial improvement in PAM image quality was achieved with only 1.4% of fully sampled pixels.
• The DIP approach significantly outperformed traditional interpolation methods.
• DIP demonstrated competitive performance compared to pre-trained supervised DL models.

Conclusions:

• Deep image prior (DIP) offers an efficient and effective solution for enhancing undersampled PAM images.
• DIP's ability to work without pre-training or ground truth makes it highly adaptable.
• The proposed method shows promise for various high-speed, undersampling imaging modalities.