Counterfactual condition diffusion with continuous prior adaptive correction for anomaly detection in multimodal brain MRI

Pixel-level prediction of early lesions is important for disease treatment and saving patients’ lives. Owing to the heterogeneity of pathological brain structures and the complexity of brain imaging, automatic anomaly detection in brain MRIs is one of the most challenging tasks in medical imaging. Recent studies have considered it as an unsupervised issue of healthy distributions. However, it is generally insufficient for modeling a normal distribution, mining abnormal data information, and considering common and complementary knowledge of multimodal data. The motivation of this study is to consider the potential performance gains of a diffusion model utilizing multimodal features for brain anomaly detection. The specific novelty of the proposal mainly includes: (1) proposing a conditional-guided diffusion model that trains healthy and abnormal images with superior performance to the unsupervised model; (2) converting pixel-level lesion prediction as a label-level counterfactual estimate and reconstructing healthy domain images with minimal intervention; (3) constructing a histogram-driven adaptive prior correction module, which enhances the utilization of multimodal information; and (4) proposing enhanced receptive field and attention modules that improve spatial perceptual loss during iterative denoising sampling. The proposed Ano-cDiff exhibited high performance in downstream brain glioma segmentation tasks. This resulted in an 8.06% increase in precision, a 0.2% increase in specificity, and a 2.58% increase in dice over the baseline diffusion model. In addition, it obtained optimal Dice and Specificity compared to state-of-the-art methods. Furthermore, massive ablation experiments demonstrated the effectiveness of the components in the proposed model. Our code is available at https://github.com/Snow1949/Ano-cDiff. •Condition diffusion generation model for pixel-level anomaly detection in brain MRI.•Modeling abnormal detection as binary counterfactual domain conversion issue.•Histogram-driven correction and image-level attention guidance accelerate inference.•Enhanced receptive fields solve spatial perceptual loss in iterative denoising.