Generative CT harmonization can catastrophically bias adipose tissue segmentation

In CT harmonization, CycleGAN and diffusion models caused -100% volume errors in adipose tissue segmentation, while a Gaussian filter had minimal error, showing task-specific validation trumps visual metrics.

• The study found a significant disconnect between task-agnostic image quality metrics (SSIM, PSNR) and clinically relevant segmentation performance.
• The authors attribute the generative models' failures partly to 3D slice inconsistencies and suboptimal model adaptation, not an inherent flaw of I2I translation.
• The Gaussian filter, a computationally simple method, provided the most consistent and stable harmonization for multi-organ segmentation.

Automated summary

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