Core Concepts
ContourDiff proposes a novel framework for unpaired image translation in medical imaging, focusing on preserving anatomical fidelity through contour-guided diffusion models.
Abstract
Abstract:
Unpaired image translation is crucial for medical image analysis.
Existing methods prioritize perceptual quality over anatomical fidelity.
ContourDiff introduces a novel framework to address this challenge.
Introduction:
Unpaired image translation facilitates segmentation across different modalities.
Maintaining anatomical consistency is essential but challenging due to structural biases between domains.
Method:
Contour-guided diffusion model proposed for unpaired image translation.
Anatomical contour representations used to guide the translation process.
Experiments and Results:
Evaluation on CT to MRI translation for lumbar spine and hip-and-thigh regions.
ContourDiff outperforms existing methods in maintaining anatomical structures.
Conclusion:
ContourDiff significantly improves anatomical fidelity in unpaired image translation.
Stats
著者らは、CTからMRIへの画像翻訳において、他の未対応の画像翻訳手法よりも0.126 Dice係数以上で性能が向上したことを示しています。
FIDスコアは、翻訳された画像と実際の出力ドメイン画像との間の解剖学的一貫性を信頼できるように反映していません。