In-vivo demonstration of Deep learning-based real-time photoacoustic imaging with LED arrays.
Paul, Avijit.
2023-04-19
-
Photoacoustic imaging is a biomedical imaging modality where the target object is illuminated by short-pulse (pulse width: 10-150 nanoseconds) laser light, and the less attenuated thermoelastic expansion is captured by the ultrasound transducers to reconstruct the initial acoustic pressure data into a spatial image. Traditionally, the imaging is performed with laser to achieve high signal-to-noise-ratio ... read more(SNR), but they are costly and bulky. An alternative option is to use LED-based illumination which is less costly and easily portable. But the deliverable energy outcome is in the nanojoule order which engenders low SNR images. To mitigate the issue, several frames can be averaged over time to generate high SNR image because LEDs have high pulse repetition frequency. Unfortunately, that averaging cannot be achieved in real-time. So, ultimately, to achieve high SNR imaging in real-time with a portable and affordable option, we have to use a downstream noise removal system with the LED-based system. Traditional non-learning noise removal algorithms such as Wiener, Median, BM3D, NLM, etc. are less effective in maintaining image quality and structure and they also need to have prior knowledge of noise type distribution. So, we aimed for the deep learning methods, especially the encoder-decoder-based framework where the training time is long, but we can test at real-time. In our work, we first employed an U-Net-based model where we found that it produces blurry outputs and is not noise type distribution invariant. To this end, we added a patch discriminator with the U-Net-based generator which essentially is a conditional generative adversarial network - we name it Denoising Pix2Pix (DenP2P) Both the issues of sharpness and noise type distribution invariance have been resolved with our DenP2P GAN network. One of our future directions is fine tuning the GAN architecture to capture both spatial & semantic information.
15-minute talk presented at the Tufts Graduate Student Council's 28th Graduate Student Research Symposium, April 19, 2023.read less - Paul, Avijit. "In-vivo demonstration of Deep learning-based real-time photoacoustic imaging with LED arrays." Presentation at the 28th Graduate Research Symposium, Tufts University, April 19, 2023.
- ID:
- 76537g39w
- To Cite:
- TARC Citation Guide EndNote
- Usage:
- Detailed Rights
