Denoising Diffusion Models for Anomaly Localization in Medical Images

This review explores anomaly localization in medical images using denoising diffusion models. After providing a brief methodological background of these models, including their application to image reconstruction and their conditioning using guidance mechanisms, we provide an overview of available datasets and evaluation metrics suitable for their application to anomaly localization in medical images. In this context, we discuss supervision schemes ranging from fully supervised segmentation to semi-supervised, weakly supervised, self-supervised, and unsupervised methods, and provide insights into the effectiveness and limitations of these approaches. Furthermore, we highlight open challenges in anomaly localization, including detection bias, domain shift, computational cost, and model interpretability. Our goal is to provide an overview of the current state of the art in the field, outline research gaps, and highlight the potential of diffusion models for robust anomaly localization in medical images.

💡 Research Summary

This review paper provides a comprehensive overview of recent advances in applying denoising diffusion models (DDMs) to anomaly localization in medical imaging. It begins by outlining the limitations of traditional fully supervised segmentation approaches (e.g., U‑Net) and earlier generative methods such as GANs and VAEs, especially when annotated data are scarce. The authors then introduce the theoretical foundations of diffusion models, describing the forward noising process (parameterized by a schedule of variances β₁…β_T) and the learned reverse denoising process p_θ, which is typically implemented with a time‑conditioned U‑Net or a diffusion transformer. Two main sampling schemes are distinguished: the stochastic DDPM and the deterministic DDIM, the latter being an ODE‑based formulation that reduces the number of required steps and improves efficiency.

The core application discussed is reconstruction‑based anomaly detection. An input image x_P is corrupted with a controllable amount of noise L, then a diffusion model trained to generate healthy samples produces a pseudo‑healthy reconstruction \hat{x}_H. The pixel‑wise anomaly map is obtained by the absolute difference |x_P − \hat{x}_H|. The noise level L directly trades off sensitivity and specificity: higher L enables stronger alteration of pathological regions but may also degrade normal tissue fidelity.

A substantial portion of the paper is devoted to guidance mechanisms that condition the diffusion process on task‑specific information. Four categories are compared:

1. Concatenation Guidance – the conditioning image is concatenated as extra channels at each denoising step. It is simple, stable, and applicable to any paired image‑to‑image task, but requires paired data.
2. Gradient Guidance – an auxiliary classifier C is trained on noisy inputs; its gradient ∇_{x_t} log C is injected into the denoising update. This plug‑and‑play approach offers flexibility but can be unstable and inherits any bias present in the classifier.
3. Classifier‑Free Guidance – the diffusion model itself is trained with an optional class label c; during sampling, c can be set to a desired class or omitted (c = ∅). This yields robust gradients with a single model, yet necessitates retraining for each new class.
4. Implicit Guidance – the model leverages the input image itself through self‑supervised masking, patch‑based conditioning, or Bernoulli diffusion. No external labels are needed, but careful design of the masking and scoring scheme is crucial. Table 1 in the paper summarizes the advantages and disadvantages of each strategy.

The review then surveys publicly available datasets used for diffusion‑based anomaly localization, covering multiple modalities (brain MRI, CT, chest X‑ray, retinal OCT) and anatomical regions. Prominent examples include BraTS, ATLAS, WMH for brain tumors and white‑matter lesions, as well as CROMIS for CT stroke imaging. Datasets differ in annotation granularity (pixel‑wise masks, bounding boxes, image‑level tags), which directly influences the choice of supervision regime.

Supervision schemes are categorized as follows:

• Fully supervised segmentation – pixel‑wise labels drive a direct U‑Net style training, achieving high accuracy at the cost of extensive annotation effort.
• Semi‑supervised learning – a small labeled subset is combined with a large unlabeled pool using pseudo‑labeling or consistency regularization, reducing annotation burden while retaining reasonable performance.
• Weakly supervised learning – only image‑level or coarse annotations are available; methods typically generate class activation maps (CAMs) that are refined by diffusion models.
• Self‑supervised learning – the diffusion model itself provides the supervisory signal (e.g., predicting added noise, masking anomalous patches).
• Unsupervised learning – models are trained exclusively on normal data; anomalies are detected as reconstruction errors when the model is forced to generate a healthy counterpart.

Each regime presents a distinct trade‑off between label cost, generalization to unseen pathologies, and clinical feasibility.

The authors conclude by highlighting four major open challenges:

1. Detection bias and domain shift – models trained on a specific scanner or population may under‑perform on data from other institutions, leading to systematic false positives/negatives.
2. Computational cost – high‑resolution 3D volumes demand substantial memory and inference time; current works often resort to 2D slice‑wise processing or reduced resolutions.
3. Model interpretability – diffusion processes are inherently black‑box; clinicians need transparent explanations, prompting research into step‑wise noise visualizations and attention‑based explanations.
4. Clinical validation and integration – rigorous prospective studies, regulatory compliance, and seamless integration into radiology workflows remain largely unexplored.

Potential solutions discussed include domain adaptation techniques, meta‑learning for rapid fine‑tuning, lightweight diffusion architectures, and hybrid models that combine diffusion priors with explicit anatomical constraints.

Overall, the paper positions denoising diffusion models as a powerful, flexible tool for medical anomaly localization, capable of operating across a spectrum of supervision levels. However, realizing their clinical impact will require addressing data heterogeneity, computational efficiency, and explainability through interdisciplinary research.