Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diffusion Models with Implicit Guidance for Medical Anomaly Detection

Cosmin I. Bercea, Benedikt Wiestler, Daniel Rueckert, Julia A. Schnabel

TL;DR

Comparative evaluations show that THOR surpasses existing diffusion-based methods in detecting and segmenting anomalies in brain MRIs and wrist X-rays and surpasses existing diffusion-based methods in detecting and segmenting anomalies in brain MRIs and wrist X-rays.

Abstract

Diffusion models have advanced unsupervised anomaly detection by improving the transformation of pathological images into pseudo-healthy equivalents. Nonetheless, standard approaches may compromise critical information during pathology removal, leading to restorations that do not align with unaffected regions in the original scans. Such discrepancies can inadvertently increase false positive rates and reduce specificity, complicating radiological evaluations. This paper introduces Temporal Harmonization for Optimal Restoration (THOR), which refines the de-noising process by integrating implicit guidance through temporal anomaly maps. THOR aims to preserve the integrity of healthy tissue in areas unaffected by pathology. Comparative evaluations show that THOR surpasses existing diffusion-based methods in detecting and segmenting anomalies in brain MRIs and wrist X-rays. Code: https://github.com/ci-ber/THOR_DDPM.

Diffusion Models with Implicit Guidance for Medical Anomaly Detection

TL;DR

Comparative evaluations show that THOR surpasses existing diffusion-based methods in detecting and segmenting anomalies in brain MRIs and wrist X-rays and surpasses existing diffusion-based methods in detecting and segmenting anomalies in brain MRIs and wrist X-rays.

Abstract

Diffusion models have advanced unsupervised anomaly detection by improving the transformation of pathological images into pseudo-healthy equivalents. Nonetheless, standard approaches may compromise critical information during pathology removal, leading to restorations that do not align with unaffected regions in the original scans. Such discrepancies can inadvertently increase false positive rates and reduce specificity, complicating radiological evaluations. This paper introduces Temporal Harmonization for Optimal Restoration (THOR), which refines the de-noising process by integrating implicit guidance through temporal anomaly maps. THOR aims to preserve the integrity of healthy tissue in areas unaffected by pathology. Comparative evaluations show that THOR surpasses existing diffusion-based methods in detecting and segmenting anomalies in brain MRIs and wrist X-rays. Code: https://github.com/ci-ber/THOR_DDPM.
Paper Structure (10 sections, 5 equations, 7 figures, 2 tables)

This paper contains 10 sections, 5 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Anomaly Detection Setup
  4. Denoising Diffusion Probabilistic Models (DDPMs)
  5. Method: THOR
  6. Experiments and Results
  7. Ischemic Stroke Lesion Segmentation in Brain MRI
  8. Anomaly Localization in Pediatric Wrist X-rays
  9. Discussions and Conclusion
  10. Appendix

Figures (7)

  • Figure 2: The top row displays the traditional DDPM de-noising sequence, where noise is progressively reduced to clarify image features. In contrast, the middle row showcases THOR: starting with equivalent high noise levels and then strategically applying unsupervised temporal anomaly masks at key intervals (indicated by orange borders) to 'harmonize' the image. This 'harmonization' process selectively refines the image by maintaining normal tissue integrity while attenuating anomalies. The bottom row shows the reverse process with anomalies becoming increasingly apparent as noise is reversed, culminating in the ground truth (GT) image where the anomaly is clearly delineated.
  • Figure 3: Anomaly detection in brain MRI scans processed by THOR using Gaussian (G) and Simplex (S) noise. From left to right, the lesions increase in size, with the smallest representing a challenging case.
  • Figure 4: Noise Level Ablation.THOR outperforms the diffusion counterparts under both Gaussian and Simplex noise types across different noise levels $T$.
  • Figure 5: Anomaly detection in pediatric wrist X-rays processed by THOR using Gaussian noise. False positives arise from unannotated non-pathological changes like unnatural bone positions following fractures or the presence of casts.
  • Figure 6: Qualitative assessment of different diffusion-based models in Brain MRI. THOR refines the performance of both DDPM (Gaussian) and AnoDDPM (Simplex), resulting in more accurate reconstructions and enhanced segmentations.
  • ...and 2 more figures