Automated three-dimensional image registration for longitudinal photoacoustic imaging

Area of Science:

• Biomedical Imaging
• Medical Physics
• Computational Imaging

Background:

• Photoacoustic tomography (PAT) shows promise for breast cancer monitoring, but repositioning variations cause geometric misalignment.
• These misalignments complicate the differentiation between true biological changes and imaging artifacts, hindering accurate disease assessment.

Purpose of the Study:

• To develop a 3D image registration framework for aligning repeated PAT volumetric images affected by repositioning.
• To address challenges like misalignment, altered light fluence, and varying fields-of-view in longitudinal PAT studies.

Main Methods:

• A coordinate-based neural network (MUVINN-reg) was developed to represent displacement fields between PAT images.
• A loss function incorporating normalized cross-correlation and multi-scale Frangi vesselness was implemented.
• The framework was validated on longitudinal PAT data from healthy volunteers, including tests under intentional mispositioning and varying breast support.

Main Results:

• MUVINN-reg demonstrated excellent performance in co-registering repeated PAT image pairs.
• The framework proved robust against image intensity shifts and field-of-view variations.
• Accurate vessel alignment was achieved at depths exceeding 4 cm.

Conclusions:

• The developed framework enables quantitative and reproducible monitoring using PAT.
• This advancement is crucial for reliable tracking of disease progression and treatment response in breast cancer.