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

TL;DR

This work addresses brain MRI anomaly segmentation under a fully weakly-supervised setting, removing the need for pixel-level labels during hyperparameter tuning. It introduces AnoFPDM, which exploits the forward diffusion process with a fixed guidance strength and dynamically selects per-input noise scale and threshold, aided by aggregation of sub-anomaly maps into a stronger aggregated anomaly map. The approach demonstrates state-of-the-art performance among weakly-supervised methods on BraTS21 and ATLAS v2.0, while providing robust per-image hyperparameter adaptation and improved signal strength for anomalous regions. Its practical significance lies in reducing labeling costs and biases, enabling scalable deployment in clinical workflows while maintaining high segmentation accuracy.

Abstract

Weakly-supervised diffusion models (DMs) in anomaly segmentation, leveraging image-level labels, have attracted significant attention for their superior performance compared to unsupervised methods. It eliminates the need for pixel-level labels in training, offering a more cost-effective alternative to supervised methods. However, existing methods are not fully weakly-supervised because they heavily rely on costly pixel-level labels for hyperparameter tuning in inference. To tackle this challenge, we introduce Anomaly Segmentation with Forward Process of Diffusion Models (AnoFPDM), a fully weakly-supervised framework that operates without the need of pixel-level labels. Leveraging the unguided forward process as a reference for the guided forward process, we select hyperparameters such as the noise scale, the threshold for segmentation and the guidance strength. We aggregate anomaly maps from guided forward process, enhancing the signal strength of anomalous regions. Remarkably, our proposed method outperforms recent state-of-the-art weakly-supervised approaches, even without utilizing pixel-level labels.

Figures (6)

• Figure 1: The diagram of the proposed method. Given the selected fixed guidance strength $w$, the noise scale $t_e$ and threshold $Q^*$ are dynamically selected for each input $\boldsymbol{x}_0$ according to the divergence $M_{t_e}$ between HFP and UFP calculated in each forward step. $\tilde{\boldsymbol{x}}_{0,t}^{h}$ and $\tilde{\boldsymbol{x}}_{0,t}^{\emptyset}$ are HGP and UGP respectively. The SAMs are aggregated to enhance the signal strength of anomalous regions.
• Figure 2: $MSE^h_t$ and $MSE_t^\emptyset$ for (a) healthy samples and (b) unhealthy samples. (c) Average change in the magnitude of SAMs in healthy and anomalous regions, alongside their ratio and divergence $M_t$ between HGP and UGP. (d) Relationship between the size of the anomalous region and the maximal divergence $M_{t_e}$.
• Figure 3: The gradient of the log-likelihood of the implicit classifier (a) and SAMs (b) at the selected steps from BraTS21 dataset
• Figure 4: Qualitative Comparison of Anomaly Maps and Segmentation. (a) From the BraTS21 dataset and (b) from the ATLAS v2.0 dataset. The first column displays the original input images, and the second column shows the corresponding ground truth for anomaly segmentation. Subsequent columns present the anomaly maps and segmentation results obtained using our method, AnoFPDM with the DDIM setting, alongside those from the second and third best comparative methods. Each row represents a different sample.
• Figure 5: Analysis of the impact of hyperparameter variations on model performance. (a) Variation in classification accuracy as a function of guidance strength $w$. (b) Corresponding changes in DICE score with different guidance stength $w$. (c) Effect of the end step $t_e$ on DICE score, using the selected guidance strength. (d) Comparison of DICE scores using a fixed threshold.