DL reconstruction for photoacoustic CT fails to recover breast lesions in synthetic benchmark study

A benchmarking framework with synthetic breast datasets and task-based image quality measures revealed that deep learning reconstruction methods for photoacoustic computed tomography often fail to accurately recover lesions, despite performing well on traditional metrics.

• The framework provides over 10,000 synthetic 2D breast objects with clinically relevant lesions and paired measurements at varying complexity.
• DL methods succeeded on traditional image quality metrics but frequently failed at lesion recovery, motivating adoption of task-based assessments.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.