Masked cross-domain self-supervised deep learning framework for photoacoustic computed tomography reconstruction

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Accurate photoacoustic computed tomography (PACT) image reconstruction is vital but often requires extensive data.
• Supervised deep learning methods for PACT demand high-quality ground truth labels, which are costly to obtain.
• Limited measurements in PACT lead to trade-offs between reconstruction cost and performance.

Purpose of the Study:

• To develop a novel self-supervised reconstruction strategy for PACT that eliminates the need for ground truth labels.
• To enable efficient and accurate PACT image reconstruction from limited measurements.
• To reduce the computational cost and improve the performance of PACT imaging.

Main Methods:

• A masked cross-domain self-supervised (CDSS) reconstruction strategy was proposed and implemented in a model-based form.
• Self-supervision was utilized to enforce consistency between image and measurement domains across partitioned PA data.
• Random masking of a substantial proportion (e.g., 80%) of measurement channels was employed.

Main Results:

• The CDSS approach effectively reduces pseudo-solutions and minimizes reconstruction error using fewer PA measurements.
• Experimental results on in-vivo mouse PACT data demonstrated the framework's potential.
• Achieved a structural similarity index (SSIM) of 0.87 with only 13 channels, outperforming a 16-channel supervised method (0.77 SSIM).

Conclusions:

• The proposed masked CDSS framework offers a viable solution for accurate PACT reconstruction with limited data.
• This self-supervised method significantly improves reconstruction quality and efficiency, particularly in sparse measurement scenarios.
• The end-to-end deployable method enhances PACT's versatility and applicability in cost-sensitive settings.