A New Deep Learning Network for Mitigating Limited-view and Under-sampling Artifacts in Ring-shaped Photoacoustic
Huijuan Zhang1, Hongyu Li1, Nikhila Nyayapathi1
1Department of Biomedical Engineering, University at Buffalo, The State University of New York, Buffalo, New York, 14260, United States.
Summary
Photoacoustic tomography (PAT) imaging quality is improved using a new deep learning method, RADL-net. This convolutional neural network effectively reduces artifacts from sparse transducer arrays, outperforming traditional methods.
Area of Science:
- Biomedical Imaging
- Medical Physics
- Optical Imaging
Background:
- Photoacoustic tomography (PAT) offers high-resolution optical absorption imaging in tissues.
- Ring-shaped transducer arrays are common in PAT but often sparse due to cost.
- Sparse arrays cause limited-view problems and under-sampling artifacts in reconstructed images.
Purpose of the Study:
- To develop a deep learning approach to overcome limited-view and under-sampling artifacts in PAT.
- To introduce the ring-array deep learning network (RADL-net) for enhanced PAT image reconstruction.
Main Methods:
- Designed a convolutional neural network (CNN) named RADL-net.
- Trained and validated RADL-net on a three-quarter ring transducer array.
- Evaluated performance using numerical simulations, phantom imaging, and in vivo studies.
Main Results:
- RADL-net successfully eliminated limited-view and under-sampling artifacts in PAT images.
- Reconstructed images showed significantly improved quality with the RADL-net method.
- The deep learning approach outperformed the conventional compressed sensing (CS) algorithm.
Conclusions:
- RADL-net offers a superior solution for artifact reduction in PAT using sparse ring transducer arrays.
- Deep learning integration significantly enhances the diagnostic potential of PAT imaging.
- The method shows promise for clinical applications requiring high-quality cross-sectional imaging.
