LED-based Photoacoustic image post-processing in real time with deep learning network.
Paul, Avijit.
Mallidi, Srivalleesha.
2022
-
Traditional Photoacoustics (PA) imaging, a non-invasive, label-free, and hybrid imaging modality, provides reasonable optical contrast along with satisfactory depth resolution, but at the cost of bulky, expensive, and low pulse repetition frequency (PRF) laser illumination. With the benefits of small footprint, portability, and high PRF, LED-based PA imaging systems are gaining a foothold to engage ... read morein the clinical translation, but they also have some shortcomings like non-tunability disabling the spectroscopy functionality, low fluence delivery engendering high noise, etc. Even though the fluence compensation is achieved by the high frame-averaging, the exploration of functional dynamics imaging at real time is impeded which might affect clinical applicability. The main goal of this research is to build a generalized deep learning software framework with U-net which not only denoises the low-frame averaged images but also increases the SNR and CNR, all apparently, at real time. The deep network was trained in Keras on top of TensorFlow in Google collab with the low (30 Hz with 128 frames) and high frame-averaging (0.15 Hz with 25600 frames) PA images of metal plates captured by the Acoustic X LED with 7 MHz central frequency and 850nm wavelength, validated with in-class sample data and then tested with out-of-class in-vitro graphite ink immersed in TiO2-Gelatin solution and in-vivo subcutaneous human pancreatic adenocarcinoma xenograft mouse model. The less-training overhead deep learning architecture is able to reduce the background noise by more than 8 times while retaining acceptable SNR, PSNR, and CNR significantly higher (p-value < 0.01) than the low-averaged inputs. Additionally, the network is highly robust to different noise profiles making it a noise invariant system. Using this particular model, more different tumor biology needs to be tested in the future along with supplementing the network with transfer learning for further generalizations.
Poster presented by Avijit Paul at the Tufts Graduate Student Council's 27th Graduate Student Research Symposium, April 20, 2022.read less - Paul, Avijit, Srivalleesha Mallidi. "LED-based Photoacoustic image post-processing in real time with deep learning network?" Poster presentation at the 27th Graduate Research Symposium, Tufts University, April 20, 2022.
- ID:
- c821gz951
- To Cite:
- TARC Citation Guide EndNote
- Usage:
- Detailed Rights
