Volumetric Directional Diffusion: Anchoring Uncertainty Quantification in Anatomical Consensus for Ambiguous Medical Image Segmentation

2026-03-04

AI Summary

This work addresses the challenges of segmenting ambiguous lesions in medical images, where high inter-observer variability and the overconfidence of conventional models—stemming from their neglect of cognitive uncertainty—hinder reliable clinical deployment. To this end, we propose a novel diffusion-based approach that incorporates a deterministic anatomical consensus prior to iteratively guide the generation of 3D boundary residual fields. This framework explicitly models the geometric variations inherent in expert annotations while preserving topological integrity. By anchoring the generative trajectory to the anatomical consensus through a volumetric directional diffusion mechanism, our method effectively prevents structural discontinuities and anatomical hallucinations, achieving a breakthrough balance between fidelity and diversity. Evaluated on the LIDC-IDRI, KiTS21, and ISBI 2015 datasets, the proposed method matches near-deterministic upper-bound segmentation accuracy while significantly advancing uncertainty quantification, attaining state-of-the-art performance in GED and CI metrics.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Security and Privacy: Data transparency and provenance

Abstract

Equivocal 3D lesion segmentation exhibits high inter-observer variability. Conventional deterministic models ignore this aleatoric uncertainty, producing over-confident masks that obscure clinical risks. Conversely, while generative methods (e.g., standard diffusion) capture sample diversity, recovering complex topology from pure noise frequently leads to severe structural fractures and out-of-distribution anatomical hallucinations. To resolve this fidelity-diversity trade-off, we propose Volumetric Directional Diffusion (VDD). Unlike standard diffusion models that denoise isotropic Gaussian noise, VDD mathematically anchors the generative trajectory to a deterministic consensus prior. By restricting the generative search space to iteratively predict a 3D boundary residual field, VDD accurately explores the fine-grained geometric variations inherent in expert disagreements without risking topological collapse. Extensive validation on three multi-rater datasets (LIDC-IDRI, KiTS21, and ISBI 2015) demonstrates that VDD achieves state-of-the-art uncertainty quantification (significantly improving GED and CI) while remaining highly competitive in segmentation accuracy against deterministic upper bounds. Ultimately, VDD provides clinicians with anatomically coherent uncertainty maps, enabling safer decision-making and mitigating risks in downstream tasks (e.g., radiotherapy planning or surgical margin assessment).

Problem

• uncertainty quantification
• medical image segmentation
• inter-observer variability
• anatomical plausibility
• 3D lesion segmentation

Innovation

• Volumetric Directional Diffusion
• Uncertainty Quantification
• Anatomical Consensus
• Medical Image Segmentation
• Boundary Residual Field