Source-linked AI summary
Image-Conditioned Diffusion Models for Quality Assurance of Organ-at-Risk Segmentations in Radiotherapy
Clea Dronne, Catharine H Clark, Xavier Loizeau, Elizabeth Miles, Peter Hoskin, Jamie R McClelland
TL;DR
Reviewing organ-at-risk segmentations is important but time-consuming and subjective, motivating model-agnostic normative QA. The paper compares a VAE with an image-conditioned segmentation diffusion model, finding that diffusion reconstruction more consistently localised subtle simulated boundary errors using regional DTA.
Problem
Manual review of organ-at-risk segmentations is time-consuming and subjective, while QA methods should assess manual, automated, and edited segmentations using patient-specific anatomical context.
Method
The study compares VAE and image-conditioned diffusion normative models that reconstruct reviewed segmentations from CT-conditioned, clinically plausible image–segmentation distributions.
Results
The diffusion model more consistently detected and localised simulated boundary errors than the VAE, particularly when DTA was evaluated within affected regions.
Takeaways & Limitations
Image-conditioned diffusion reconstruction is a promising framework for localised, anatomy-aware segmentation QA that supports rather than replaces manual clinical review.
Takeaways & Limitations
Evaluation used simulated perturbations on variable RADCURE segmentations, so it may not capture the full variability of real clinical errors; future work will assess real errors and additional OARs.
Abstract
from arXiv · showhide
Accurate organ-at-risk segmentation is essential for radiotherapy planning, but reviewing segmentations is time-consuming and subjective. We investigate normative modelling for segmentation error detection in head-and-neck CT, comparing a VAE framework with an image-conditioned segmentation diffusion model. Models were evaluated on RADCURE brainstem and spinal cord segmentations using simulated boundary and width perturbations. Error detection was assessed using the Dice similarity coefficient and the Distance to Agreement (DTA) between the input and reconstructed segmentations. While both models detected some simulated errors, regional DTA showed that the diffusion model localised subtle boundary errors more consistently. These results support image-conditioned diffusion reconstruction as a promising framework for localised, anatomy-aware segmentation QA.
1 Introduction
The paper addresses time-consuming, subjective review of organ-at-risk segmentations by evaluating normative modelling with an image-conditioned diffusion model against a VAE baseline for simulated error detection and localisation.
- Accurate OAR delineation is essential for radiotherapy planning and clinical-trial QA because segmentation errors can reduce tumour dose coverage and increase healthy-tissue toxicity.
- Clinical OAR segmentations require review regardless of whether they are manual, automated, or produced through hybrid AI-assisted workflows.
- Post-hoc automated QA aims to identify cases or regions needing closer inspection while remaining model-agnostic across segmentation-generation workflows.
- Shape-only methods and summary metrics can miss patient-specific or spatially localised segmentation errors, motivating the use of anatomical image context.
- Normative modelling reconstructs a reviewed segmentation from the learned distribution of clinically acceptable image–segmentation pairs, using discrepancies as a QA signal.
- The study compares an established VAE QA approach with a new image-conditioned segmentation diffusion model for detecting and localising simulated OAR errors in 3D patches.
2 Methodology
Both models reconstruct reviewed segmentations using CT anatomy, while the proposed diffusion model corrupts only the segmentation and denoises it with fixed CT conditioning.
- Both approaches learn anatomically plausible segmentations from clinically approved image–segmentation pairs within a normative modelling framework.
- The VAE receives CT and binary OAR segmentation as multi-channel inputs and reconstructs both through a probabilistic latent space.
- The diffusion model reconstructs a noise-corrupted segmentation while using the uncorrupted CT image as anatomical conditioning information.
- Given CT image x0 and input segmentation y0, the framework reconstructs ŷ0 and performs QA by comparing the input and reconstruction.
- A Bernoulli diffusion process progressively corrupts segmentation voxels, with the resulting noisy mask denoised by a 3D U-Net conditioned on multi-scale CT features.
- The model predicts voxel-wise Bernoulli flip-noise logits and is trained with a hybrid variational-lower-bound and binary-cross-entropy objective.
- At inference, the segmentation is partially corrupted to timestep τ < T, denoised with deterministic DDIM sampling, averaged across high-corruption timesteps, and binarised at 0.5.
3 Experiments
Experiments compared VAE and image-conditioned diffusion reconstruction for detecting and localising simulated brainstem and spinal cord segmentation errors. Regional DTA showed more consistent localisation for the diffusion model, although sensitivity varied by perturbation.
- Dataset and models: Experiments used RADCURE head-and-neck CT data, focusing on brainstem and spinal cord segmentations with separate organ-specific models.Training used 2,093 brainstem and 1,993 spinal cord cases, with fixed validation and testing splits.
- Evaluation procedure: QA signals were derived from voxel-wise DTA maps measuring disagreement between each input segmentation and its reconstruction.DTA was computed over the foreground union, with zero distance for shared voxels and nearest-agreement distances for discrepant voxels.
- Evaluation procedure: Simulated errors included superior and inferior brainstem boundary shifts, plus superior boundary shifts and transverse widening for the spinal cord.The evaluation followed the published VAE perturbation protocol and compared DSC and DTA against corresponding nonperturbed segmentations.
- Results: Figure 2 showed low DTA without errors and higher responses in perturbed regions, with the diffusion model producing clearer localisation across the illustrated cases.The VAE response was weak for the −3 ↓ brainstem perturbation and less localised in the other examples.
- Results: Regional DTA detected subtle brainstem perturbations more often with diffusion than VAE, while neither model significantly detected +1 ↑ or +2 ↑ superior extensions.The diffusion model significantly detected −2 ↑, −1 ↑, −1 ↓, and +1 ↓ brainstem perturbations; spinal-cord results included +3 ↑, −1 ↑, and +2 ↔, with more consistent detection and localisation than VAE.
4 Discussion
The diffusion model more consistently detected and localised simulated boundary errors than the VAE, while preserving patient-specific anatomical context during reconstruction. The approach remains intended to support manual review, with runtime and dataset variability limiting practical interpretation.
- The diffusion model more consistently detected and localised simulated boundary errors than the VAE, particularly using regional DTA.
- The VAE’s shared latent representation can blur CT reconstructions, reducing anatomical detail and sensitivity to subtle segmentation errors.
- Keeping the CT image uncorrupted during segmentation denoising preserved anatomical context and produced DTA maps that more clearly highlighted error regions.
- RADCURE segmentations vary substantially, particularly at soft-tissue boundaries, because many predated clear delineation guidelines.
- Simulated perturbations may not capture the full variability of real clinical segmentation errors, motivating evaluation on real errors and additional organs.
- Diffusion reconstruction required 4.6s per image per organ across four timesteps, compared with 0.5s for the VAE, but may suit offline QA workflows.