AnoFPDM: Anomaly Detection with Forward Process of Diffusion Models for Brain MRI

Yiming Che, Fazle Rafsani, Jay Shah, Md Mahfuzur Rahman Siddiquee, Teresa Wu; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 1113-1122

Weakly-supervised diffusion models in anomaly segmentation which leverage image-level labels and bypass the need for pixel-level labels during training have shown superior performance over unsupervised methods offering a cost-effective alternative. Traditional methods that rely on iterative image reconstruction are not fully weakly-supervised due to their dependence on costly pixel-level labels for hyperparameter tuning in inference. To address this issue we introduce Anomaly Detection with Forward Process of Diffusion Models (AnoFPDM) a fully weakly-supervised framework that operates without image reconstruction and eliminates the need for pixel-level labels in hyperparameter tuning. By using the unguided forward process as a reference AnoFPDM dynamically selects hyperparameters such as noise scale and segmentation threshold for each input. We improve anomaly segmentation by aggregating anomaly maps from each step of the guided forward process which strengthens the signal of anomalous regions in the aggregated anomaly map. Our framework demonstrates competitive performance on the BraTS21 and ATLAS v2.0 datasets. Code is available at https://github.com/SoloChe/AnoFPDM.