Anatomically constrained liver CT anomaly detection using healthy priors with diffusion-based inpainting.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Detecting subtle liver lesions on CT scans is difficult due to variations in tumor characteristics and imaging protocols.
• Fully supervised segmentation methods require extensive manual annotation, limiting scalability.
• Existing anomaly detection methods lack anatomical constraints, leading to boundary instability and sensitivity to image variations.

Purpose of the Study:

• To develop an anatomically constrained, four-stage pipeline for anomaly detection in liver CT scans.
• To improve the accuracy and generalizability of liver lesion detection without extensive manual annotation.
• To create a data-efficient method for localizing liver lesions for CT triage and prioritization.

Main Methods:

• A denoising diffusion probabilistic model (DDPM) was trained on healthy liver CT slices to establish a normal appearance prior.
• Diffusion-based inpainting generated pseudo-normal liver images within an automatically segmented liver mask.
• An encoder-decoder model reconstructed healthy liver tissue using paired original and inpainted images.
• A liver-scoped difference map between original and reconstructed images served as the anomaly score.

Main Results:

• The method achieved a Dice score of 0.596, IoU of 0.482, and AUROC of 0.861 on abnormal CT scans.
• Performance improved with lesion size, reaching a Dice score of 0.796 for the largest quartile of lesions.
• The approach demonstrated data-efficient liver lesion localization, outperforming methods without anatomical constraints.
• The 95th percentile Hausdorff distance was 80.5 pixels.

Conclusions:

• Anatomically constrained anomaly detection with a stable healthy prior enhances liver lesion localization in CT scans.
• The proposed four-stage pipeline offers a data-efficient and scalable solution for detecting subtle liver lesions.
• This method shows promise for improving CT triage and prioritization in routine clinical practice.