💻 computer science

Image-Conditioned Diffusion Models for Quality Assurance of Organ-at-Risk Segmentations in Radiotherapy

This paper demonstrates that image-conditioned diffusion models outperform VAEs in detecting and localizing subtle segmentation errors in head-and-neck radiotherapy planning, offering a promising framework for automated quality assurance of organ-at-risk delineations.

Original authors: Clea Dronne, Catharine H Clark, Xavier Loizeau, Elizabeth Miles, Peter Hoskin, Jamie R McClelland

Published 2026-08-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Clea Dronne, Catharine H Clark, Xavier Loizeau, Elizabeth Miles, Peter Hoskin, Jamie R McClelland

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Radiation therapy is a powerful tool for treating cancer, using high-energy beams to destroy tumors while sparing the healthy tissue around them. For this treatment to work safely, doctors must first create a precise map of the patient's internal anatomy, outlining the tumor and the nearby organs that could be damaged by the radiation. These outlines, known as segmentations, are critical; if they are even slightly wrong, the treatment might miss the cancer or harm healthy organs. Currently, creating and checking these maps is a slow, manual job that relies on human experts to review every single case. Because this process is so time-consuming and depends on the individual judgment of the reviewer, there is a growing need for computer systems that can help spot mistakes quickly and reliably, acting as a second pair of eyes to ensure patient safety.

In a recent study, researchers explored a new way to build these computer assistants using a type of artificial intelligence called a diffusion model. The team focused on the head and neck, where the brainstem and spinal cord are vital structures that must be protected. They compared their new approach against an older method that had been used for similar tasks. The goal was to see if the new system could not only tell that a map was wrong but also pinpoint exactly where the error was located, such as a boundary that was drawn too high or too low. To test this, the researchers took a large collection of real medical scans and the corresponding organ maps that had already been approved for use. They then deliberately introduced small, controlled mistakes into these maps, simulating the kinds of errors that might happen in a real clinic, such as shifting a boundary line or making an organ look slightly wider than it should be.

The researchers trained two different computer systems to learn what a correct map should look like for a specific patient's anatomy. The first system was a standard model that had been used in previous studies. The second, newer system was designed to look at the patient's scan and use that visual information to guide its thinking, much like a draftsman who constantly refers to a photograph while drawing a sketch. When the computer was given a map with a mistake, it tried to "reconstruct" what the correct map should have been based on the scan. If the computer's corrected version looked very different from the submitted map, the system flagged that area as potentially erroneous. The researchers then measured how well each system could detect these simulated errors and how accurately it could highlight the specific spot where the mistake occurred.

The results showed that while both systems could identify that something was wrong, the newer diffusion model was significantly better at finding the exact location of the error. When the researchers looked at the specific regions where they had introduced mistakes, the new model consistently produced clear signals that pointed directly to the problem area. In contrast, the older system often missed subtle errors or failed to localize them precisely. For example, when a boundary was shifted by just a small amount, the new model was able to detect the discrepancy and highlight it, whereas the older model often treated the map as acceptable. This ability to focus on the specific region of concern is crucial for clinical use, as it allows a human reviewer to look directly at the suspicious area rather than having to scan the entire image to find the fault.

The study suggests that this new approach offers a promising way to improve the quality control of radiation therapy planning. By using the patient's own scan as a guide, the new model preserves the unique anatomical details needed to spot subtle boundary errors that other methods might miss. The researchers noted that the older method sometimes blurred the details of the scan during its processing, which made it harder to judge the fine lines of the organ boundaries. The new method avoided this by keeping the scan sharp and using it only to inform the correction process. While the new system takes longer to run than the older one, the extra time is considered acceptable for a quality check that happens before a patient receives treatment. Ultimately, the findings indicate that this image-guided approach could become a valuable tool for helping medical teams ensure that every radiation plan is as safe and accurate as possible.

Technical Summary: Image-Conditioned Diffusion Models for QA of OAR Segmentations

Problem Statement
Accurate segmentation of Organs-at-Risk (OARs) is critical for radiotherapy treatment planning and clinical trial quality assurance (QA). However, the review of these segmentations is currently a time-consuming, subjective manual process. While automated segmentation QA methods exist, many rely on uncertainty quantification tied to specific models or use heuristic metrics that fail to capture full spatial information. The authors posit that normative modelling—learning the distribution of clinically acceptable image-segmentation pairs—offers a model-agnostic solution. In this framework, a model reconstructs a segmentation based on anatomical context; discrepancies between the input and the reconstruction serve as a signal for potential errors. While Variational Autoencoders (VAEs) have been used for this purpose, they may struggle with subtle boundary deviations due to blurred reconstructions. This work investigates whether image-conditioned Denoising Diffusion Probabilistic Models (DDPMs) can provide sharper, more spatially faithful reconstructions to improve the detection and localisation of segmentation errors.

Methodology
The study compares a baseline VAE framework with a proposed image-conditioned segmentation diffusion model. Both operate within a normative modelling paradigm using 3D patch-based inputs (64×64×64 voxels) centered on the brainstem or spinal cord from the RADCURE head-and-neck CT dataset.

VAE Baseline: The VAE takes the CT image and binary OAR segmentation as multi-channel inputs, reconstructing both from a probabilistic latent space.
Proposed Diffusion Model: This model applies a diffusion process exclusively to the segmentation mask ( $y_0$ $y_{0}$ ), while the CT image ( $x_0$ $x_{0}$ ) remains uncorrupted and serves as anatomical conditioning.
- Corruption Process: The segmentation is corrupted via a Bernoulli diffusion process where voxels are randomly flipped with increasing probability over $T=1000$ timesteps.
- Architecture: A 3D U-Net predicts the Bernoulli flip-noise logits. The uncorrupted CT image is resampled and projected via lightweight convolutional blocks at each encoder resolution to condition the segmentation features.
- Training & Inference: The model is trained using a hybrid loss combining the Bernoulli Variational Lower Bound (VLB) and Binary Cross-Entropy (BCE). During inference, the input segmentation is partially corrupted to a high timestep ( $\tau \in \{700, 800, 900, 1000\}$ ) and denoised using deterministic DDIM sampling to produce a reconstruction ( $\hat{y}_0$ ).
Evaluation Metrics: Quality Assurance is assessed by comparing the input segmentation ( $y_0$ ) and the reconstruction ( $\hat{y}_0$ ) using the Dice Similarity Coefficient (DSC) and the Distance to Agreement (DTA). Crucially, DTA maps are generated to localise errors, measuring the Euclidean distance to the nearest agreeing voxel for regions where the segmentations differ.

Key Contributions and Results
The authors evaluated both models on simulated boundary perturbations (superior/inferior shifts and transverse widening) applied to brainstem and spinal cord segmentations.

Localisation of Errors: While both models detected some errors, the diffusion model demonstrated superior capability in localising subtle boundary errors. Regional DTA analysis showed that the diffusion model produced clear, localised responses to perturbations (e.g., $-2 \uparrow$ , $-1 \uparrow$ , $-1 \downarrow$ for the brainstem; $+3 \uparrow$ , $-1 \uparrow$ , $+2 \leftrightarrow$ for the spinal cord) where the VAE often failed to show significant regional DTA increases.
Robustness to Subtle Perturbations: The diffusion model consistently detected smaller perturbations that the VAE missed, particularly in the brainstem. For instance, the VAE did not significantly detect $-2 \uparrow$ or $-1 \uparrow$ perturbations, whereas the diffusion model did.
Anatomical Fidelity: The results suggest that preserving the uncorrupted CT image as conditioning information allows the diffusion model to maintain patient-specific anatomical details better than the VAE, which reconstructs both image and segmentation from a shared latent space, potentially leading to blurred anatomical guidance.

Significance and Claims
The paper claims that image-conditioned diffusion reconstruction is a promising framework for localised, anatomy-aware segmentation QA. The primary significance lies in the model's ability to not only flag a segmentation as potentially erroneous but to direct the reviewer's attention to the specific region requiring inspection via DTA maps.

The authors maintain a modest tone regarding the scope of their findings:

The method is intended to support, not replace, manual clinical review by prioritising cases or regions for expert inspection.
The evaluation relies on simulated perturbations, which may not fully capture the variability of real clinical segmentation errors, particularly given the historical variability in the RADCURE dataset where delineation guidelines were not always clear.
A practical limitation noted is increased inference time (4.6s per reconstruction vs. 0.5s for the VAE), though the authors argue this is acceptable for offline QA workflows.

Future work is identified as necessary to validate the method on real clinical segmentation errors and to extend the approach to additional OARs.

Technical Summary: Image-Conditioned Diffusion Models for QA of OAR Segmentations

More like this