Recovering from missing data in population imaging - Cardiac MR image imputation via conditional generative...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate cardiac function assessment relies on complete Cardiac Magnetic Resonance (CMR) imaging volumes.
• Image artifacts in CMR sequences often lead to missing or unusable slices, hindering accurate quantification.
• Current methods for recovering missing CMR data are insufficient for population-level imaging studies.

Purpose of the Study:

• To develop a robust method for imputing missing cardiac short-axis slices in CMR images.
• To improve the accuracy of anatomical and functional cardiac quantification despite image artifacts.

Main Methods:

• Proposed a novel Image Imputation Generative Adversarial Network (I2-GAN) for inferring missing CMR slices.
• Utilized regression and generator networks with residual blocks and normalization layers for slice synthesis.
• Implemented a multi-scale discriminator with feature matching loss to enhance realism and performance.

Main Results:

• Achieved significant improvements over state-of-the-art methods in missing slice imputation for CMR.
• Obtained an average Structural Similarity Index Measure (SSIM) of 0.872 for imputed slices.
• Demonstrated excellent agreement between reference and imputed images for ventricular volume and mass measurements (correlation coefficients up to 0.991).

Conclusions:

• The proposed I2-GAN method robustly imputes missing CMR slices, enhancing the accuracy of cardiac quantification.
• This approach offers a promising solution for improving the reliability of CMR-based cardiac assessments in large-scale studies.
• I2-GAN facilitates more accurate population imaging by addressing the challenge of incomplete CMR data.