Deep neural network-based bandwidth enhancement of photoacoustic data

Abstract

Photoacoustic (PA) signals collected at the boundary of tissue are always band-limited. A deep neural network was proposed to enhance the bandwidth (BW) of the detected PA signal, thereby improving the quantitative accuracy of the reconstructed PA images. A least square-based deconvolution method that utilizes the Tikhonov regularization framework was used for comparison with the proposed network. The proposed method was evaluated using both numerical and experimental data. The results indicate that the proposed method was capable of enhancing the BW of the detected PA signal, which inturn improves the contrast recovery and quality of reconstructed PA images without adding any significant computational burden.

Image-guided filtering for improving photoacoustic tomographic image reconstruction

“…Note that for the data having SNR of 20 dB, the improvement is as high as 25% (results are not shown). It is also important to note that there are sophisticated methods proposed in the literature, 34,35 which could improve the reconstruction performance significantly (latest one being based on deep learning 35 ), but the computational complexity of these methods is at least one order higher than the wavelet denoising method utilized here, which has the computational complexity similar to the proposed guided filter approach.…”

“…[28][29][30] To improve the PA imaging performance, the earlier attempts included applying signal enhancement methods on the PA data collected. [31][32][33][34][35] These methods typically apply deconvolution on the raw data collected to improve recorded acoustic signal. [31][32][33][34][35] These approaches have limited utility (will also be shown with an example using the method attempted in Ref.…”

“…[31][32][33][34][35] These methods typically apply deconvolution on the raw data collected to improve recorded acoustic signal. [31][32][33][34][35] These approaches have limited utility (will also be shown with an example using the method attempted in Ref. 33) compared to postprocessing of reconstructed PA images.…”

Compensating for visibility artefacts in photoacoustic imaging with a deep learning approach providing prediction uncertainties

“…Over the past two years, a few groups started investigating deep learning applied to PA imaging for several purposes including direct reconstruction of the initial pressure [20] , handling artefacts coming from sparse data [21] , [22] , [23] , reflection artefacts removal [24] , [25] , point source localization [26] , [27] and quantitative measurements [28] , [29] . The correction of the limited bandwidth problem was also investigated on very simple objects [30] . Some of these studies [22] , [23] , [31] showed that deep learning can also reduce the limited view artefacts although results were either numerical or obtained with non-conventional imaging devices.…”

Photoacoustic imaging in the second near-infrared window: a review

“…), it undergoes a time-varying thermal expansion-relaxation process, which leads to the generation of acoustic waves (known as PA waves) in the tissue. By acquiring the generated PA waves at the tissue surface and using various reconstruction algorithms, [28][29][30][31][32][33][34][35] the absorption maps within the tissue can be obtained. By combining optical contrast and US resolution, PAI offers several advantages: (i) it is noninvasive and uses nonionizing radiation, hence it can be repeatedly used in vivo by keeping the excitation energy below the safety limit; (ii) it provides label-free imaging; (iii) it is speckle-free; (iv) it has higher penetration depth, up to ∼4 cm in vivo 36 and ∼12 cm in vitro; 37 (v) it is faster and less expensive compared to MRI, PET, x-ray CT; 38 (vi) it provides multiscale imaging, allows imaging organelles to organs with consistent contrast while keeping the same depth-toresolution ratio; 38,39 (vii) it provides direct imaging of optical absorption with 100% relative sensitivity-the sensitivity of PAI is 100 times higher than that of OCT and confocal microscopy; and (viii) it provides anatomical, functional, molecular, and kinetic information.…”