Benchmarking Deep Learning-Based Reconstruction Methods for Photoacoustic Computed Tomography with Clinically...

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) methods are advancing photoacoustic computed tomography (PACT) image reconstruction.
• Current DL evaluations lack standardized datasets and clinically relevant metrics.
• This hinders reproducible comparisons and reliable assessment of PACT advancements.

Purpose of the Study:

• To introduce a standardized benchmarking framework for DL-based acoustic inversion in PACT.
• To provide open-source synthetic datasets and evaluation strategies for PACT.
• To enable fair, reproducible, and clinically relevant comparisons of reconstruction methods.

Main Methods:

• Developed a framework with over 11,000 2D synthetic breast objects and lesions.
• Incorporated paired measurements with varying complexity.
• Integrated traditional and task-based image quality (IQ) metrics for evaluation.

Main Results:

• The framework enabled quantitative comparison of DL and physics-based methods.
• Some DL methods excelled in traditional IQ but failed to recover lesions accurately.
• Highlighted the inadequacy of traditional IQ metrics and the need for task-based assessments.

Conclusions:

• The proposed framework facilitates systematic comparisons of DL acoustic inversion methods for 2D PACT.
• It promotes reproducible, objective assessments using clinically relevant data.
• Aids in PACT method development and system optimization.