See More, Change Less: Anatomy-Aware Diffusion for Contrast Enhancement

Summary

This paper introduces SMILE, an anatomy-aware diffusion model designed for contrast enhancement in medical images without introducing common artifacts like organ distortion or false findings. SMILE selectively enhances clinically relevant regions by learning organ shapes and contrast dynamics, demonstrating superior image quality and significant improvements in cancer detection, particularly from non-contrast CT scans.

Medical Relevance

This work is highly relevant for medical imaging and diagnostics as it addresses critical limitations of current image enhancement techniques that can compromise diagnostic accuracy. By providing anatomically accurate and diagnostically meaningful enhanced images, SMILE directly supports better clinical decision-making and potentially improves the early detection of diseases like cancer.

AI Health Application

The AI application is an 'anatomy-aware diffusion model' (SMILE) that enhances medical images (specifically CT scans, multi-phase and non-contrast) by learning organ shapes and contrast dynamics. This enhancement aims to improve the visual quality of medical images, aid clinical decision-making, prevent false findings, and significantly boost the detection accuracy of conditions like cancer from non-contrast scans.

Key Points

• SMILE is an anatomy-aware diffusion model that learns true organ boundaries and contrast uptake patterns, enhancing only clinically relevant regions.
• It prevents common issues in existing enhancement models such as organ distortion, creation of false findings, and missing small tumors due to a lack of anatomical understanding.
• The model incorporates structure-aware supervision, ensuring adherence to actual anatomical contours and contrast dynamics.
• SMILE features registration-free learning, allowing it to process unaligned multi-phase CT scans directly without preprocessing alignment.
• It provides unified inference, ensuring fast and consistent enhancement across all contrast phases of a scan.
• Evaluated on six external datasets, SMILE achieved significant improvements in image quality metrics: 14.2% higher SSIM, 20.6% higher PSNR, and 50% better FID compared to existing methods.
• Clinically, it produced anatomically accurate and diagnostically meaningful images, raising the F1 score for cancer detection from non-contrast CT scans by up to 10%.

Clinical Impact

SMILE has the potential to significantly enhance diagnostic confidence and accuracy in radiology by providing clearer, anatomically faithful images that highlight clinically relevant features. Its ability to improve cancer detection from non-contrast CT could lead to earlier diagnoses, potentially reducing the need for contrast agents in some cases and streamlining workflows by offering consistent enhancement across multi-phase studies.

Limitations

The abstract does not explicitly state any limitations of the SMILE model. However, potential general limitations for such models might include performance variability across different patient populations, scanner types, or specific pathologies not covered in the six external datasets used for evaluation.

Future Directions

The abstract does not explicitly mention future research directions. However, potential avenues could include extending SMILE to other imaging modalities (e.g., MRI), evaluating its impact in prospective clinical trials, or exploring its application in automated disease grading or treatment response assessment.

Medical Domains

Keywords

Abstract

Image enhancement improves visual quality and helps reveal details that are hard to see in the original image. In medical imaging, it can support clinical decision-making, but current models often over-edit. This can distort organs, create false findings, and miss small tumors because these models do not understand anatomy or contrast dynamics. We propose SMILE, an anatomy-aware diffusion model that learns how organs are shaped and how they take up contrast. It enhances only clinically relevant regions while leaving all other areas unchanged. SMILE introduces three key ideas: (1) structure-aware supervision that follows true organ boundaries and contrast patterns; (2) registration-free learning that works directly with unaligned multi-phase CT scans; (3) unified inference that provides fast and consistent enhancement across all contrast phases. Across six external datasets, SMILE outperforms existing methods in image quality (14.2% higher SSIM, 20.6% higher PSNR, 50% better FID) and in clinical usefulness by producing anatomically accurate and diagnostically meaningful images. SMILE also improves cancer detection from non-contrast CT, raising the F1 score by up to 10 percent.