Reverse Imaging: Any-Sequence Generalization for Cardiac MRI Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Cardiac MRI segmentation models struggle with contrast variations across different imaging sequences.
• These variations stem from diverse imaging protocols, despite underlying consistent tissue properties (M0, T1, T2).

Purpose of the Study:

• To introduce Reverse Imaging, a physics-driven framework for data augmentation and domain generalization in cardiac MRI.
• To improve the robustness and generalizability of segmentation models across various cardiac MRI sequences.

Main Methods:

• Inferred tissue properties (T1, T2) by solving regularized inverse problems using a generative prior learned from mSASHA data.
• Characterized imaging sequences by information content and ill-posedness to guide iterative prior learning.
• Integrated MRI physics into posterior inference and developed a ControlNet for cine sequences.

Main Results:

• Reverse Imaging generated plausible tissue-property estimates across eight unseen cardiac MRI sequences.
• The method synthesized diverse, physically consistent contrasts and improved segmentation robustness under cross-sequence shifts.
• Demonstrated improved efficiency and spatial consistency for weak cine observations.

Conclusions:

• Reverse Imaging offers a robust solution for domain generalization in cardiac MRI segmentation.
• The physics-driven approach effectively handles contrast variations and improves model performance.
• This framework enables more reliable automated analysis of cardiac MRI data.