[Paper Review] Sinogram super-resolution and denoising convolutional neural network (SRCN) for limited data photoacoustic tomography

Navchetan Awasthi, Rohit Pardasani+3 more|arXiv (Cornell University)|Jan 17, 2020

Photoacoustic and Ultrasonic Imaging28 references18 citations

TL;DR

This paper proposes a deep learning-based sinogram super-resolution and denoising convolutional neural network (SRCN) to enhance photoacoustic tomography (PAT) image quality from limited detector data. By super-resolving and denoising sinogram data using a 7-layer CNN with residual learning, the method reduces artifacts and noise, achieving up to 41.70% lower RMSE and 6.93 dB higher PSNR compared to conventional interpolation and wavelet denoising techniques in both numerical and in-vivo experiments.

ABSTRACT

The quality of the reconstructed photoacoustic image largely depends on the amount of photoacoustic (PA) boundary data available, which in turn is proportional to the number of detectors employed. In case of limited data (owing to less number of detectors due to cost/instrumentation constraints), the reconstructed PA images suffer from artifacts and are often noisy. In this work, for the first time, a deep learning based model was developed to super resolve and denoise the photoacoustic sinogram data. The proposed method was compared with existing nearest neighbor interpolation and wavelet based denoising techniques and was shown to outperform them both in numerical and in-vivo cases. The improvement obtained in Root Mean Square Error (RMSE) and Peak Signal to Noise Ratio (PSNR) for the reconstructed PA image using the sinogram data that was super-resolved and denoised using proposed neural network based method was as high as 41.70 % and 6.93 dB respectively compared to utilizing limited sinogram data.

Motivation & Objective

• To address the degradation in photoacoustic image quality caused by limited detector data in PAT, which leads to artifacts and noise.
• To develop a data-driven deep learning model that super-resolves and denoises sinogram data to improve reconstruction fidelity.
• To outperform traditional methods such as nearest-neighbor interpolation and discrete wavelet transform (MODWT) denoising in both numerical and in-vivo settings.
• To provide a generalizable, fully data-driven framework applicable to sinogram data of any dimension for improved PAT reconstruction.

Proposed method

• A 7-layer convolutional neural network (SRCN) is trained to predict high-resolution, denoised sinograms from low-resolution, noisy sinograms obtained with 50 detectors.
• The network uses residual learning, with each convolutional layer followed by ReLU activation and batch normalization, and outputs a residual map for denoised sinogram reconstruction.
• Input sinograms are preprocessed via nearest-neighbor interpolation to upsample to 100 detector positions before being fed into the SRCN.
• The model is trained end-to-end using mean squared error (MSE) loss between predicted and ground-truth sinograms, with optimization via Adam.
• The architecture is evaluated on numerical phantoms (Shepp-Logan, Derenzo) and in-vivo rat brain data, with reconstruction quality assessed via RMSE and PSNR.
• The method is fully open-source, with code provided for dataset generation, model training, and inference in Python and MATLAB.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively super-resolve and denoise photoacoustic sinograms from limited detector data to improve image reconstruction quality?
• RQ2 How does the proposed SRCN model compare to conventional interpolation and wavelet-based denoising methods in terms of artifact reduction and noise suppression?
• RQ3 To what extent does the SRCN model generalize across different signal-to-noise ratios (SNR) and phantom configurations?
• RQ4 Can the SRCN model improve in-vivo photoacoustic image quality when applied to real biological data with limited detector acquisition?

Key findings

• The SRCN method achieved a 41.70% reduction in RMSE and a 6.93 dB improvement in PSNR for in-vivo rat brain data compared to reconstruction using only 50 detectors.
• For the 60 dB SNR numerical phantom, the SRCN reduced RMSE by 41.70% and increased PSNR by 4.26 dB compared to nearest-neighbor interpolation and MODWT denoising.
• In the Derenzo phantom, the SRCN enabled clear visualization of small-radius circles that were indistinguishable in other reconstruction methods due to aliasing artifacts.
• The SRCN outperformed both nearest-neighbor interpolation and MODWT-based denoising across all SNR levels (20, 40, 60 dB), with consistent improvements in RMSE and PSNR.
• The method successfully mitigated background noise and reduced structural artifacts in reconstructed PA images, especially in in-vivo data, despite some missing structures due to configuration mismatch.
• The SRCN demonstrated strong generalization capability on unseen data, confirming its robustness and applicability to real-world PAT scenarios with limited data.