Revisiting Deep Ensemble Uncertainty for Enhanced Medical Anomaly Detection

TL;DR

Redundancy-Aware Repulsion (RAR) is proposed, which uses a similarity kernel that remains invariant to both isotropic scaling and orthogonal transformations, explicitly promoting diversity in learners' feature space, and Dual-Space Uncertainty (DSU), which utilizes the ensemble's uncertainty in input and output spaces.

Abstract

Medical anomaly detection (AD) is crucial in pathological identification and localization. Current methods typically rely on uncertainty estimation in deep ensembles to detect anomalies, assuming that ensemble learners should agree on normal samples while exhibiting disagreement on unseen anomalies in the output space. However, these methods may suffer from inadequate disagreement on anomalies or diminished agreement on normal samples. To tackle these issues, we propose D2UE, a Diversified Dual-space Uncertainty Estimation framework for medical anomaly detection. To effectively balance agreement and disagreement for anomaly detection, we propose Redundancy-Aware Repulsion (RAR), which uses a similarity kernel that remains invariant to both isotropic scaling and orthogonal transformations, explicitly promoting diversity in learners' feature space. Moreover, to accentuate anomalous regions, we develop Dual-Space Uncertainty (DSU), which utilizes the ensemble's uncertainty in input and output spaces. In input space, we first calculate gradients of reconstruction error with respect to input images. The gradients are then integrated with reconstruction outputs to estimate uncertainty for inputs, enabling effective anomaly discrimination even when output space disagreement is minimal. We conduct a comprehensive evaluation of five medical benchmarks with different backbones. Experimental results demonstrate the superiority of our method to state-of-the-art methods and the effectiveness of each component in our framework. Our code is available at https://github.com/Rubiscol/D2UE.

Figures (4)

• Figure 1: (a): An illustration of redundancy-aware repulsion (RAR). Disagreement on anomalies is amplified between different learners’ feature spaces, while normal input converges to similar reconstructions guided by reconstruction training. (b): A t-SNE tSNE plot of feature spaces from three learners on the anomaly. Feature spaces are pushed away by RAR during training. (c): An illustration of dual-space uncertainty (DSU) in 1D regression with two learners. Utilizing output space uncertainty fails to differentiate the anomaly at the upper point. In comparison, DSU utilizes the disagreement on $\nabla_{X}{f}$ to detect such anomalies.
• Figure 2: Overview of D2UE. In the training stage, the redundancy-aware repulsion (RAR) module amplifies the diversity of different models with both isotropic and scaling invariance. In the inference stage, the dual-space uncertainty is calculated, utilizing both $f(X)$ in the output space and $\nabla_{X}\mathcal{L}$ in the input space.
• Figure 3: An illustration of neural network redundancy. Different features may output the same by weight re-scaling or spatial reordering.
• Figure 4: Visualization results. Red bounding boxes indicate abnormal regions.