Introduction
In medical imaging, anomaly detection is often framed as a One-Class Classification problem. Because diseases and lesions are rare and highly varied, models are typically trained only on "healthy" data. A common strategy to identify anomalies is the use of Deep Ensembles: if multiple models are trained on healthy data, they should theoretically agree on normal samples but disagree significantly when encountering an anomaly.
However, standard ensembles often suffer from simplicity bias, where different models learn redundant features. This redundancy leads to minimal disagreement on anomalies, rendering them invisible to the system. To address this, the authors propose D2UE (Diversified Dual-space Uncertainty Estimation), a framework that enforces diversity among ensemble members while maintaining their accuracy on normal data.
Methodology
The D2UE framework introduces two primary innovations to break model redundancy and enhance detection sensitivity:
1. Redundancy-Aware Repulsion (RAR)
During the training phase, models are trained sequentially. Each new model is forced to be "repulsed" by the feature representations of its predecessors. The core of this repulsion is the use of Centered Kernel Alignment (CKA). Unlike simple distance metrics, CKA is invariant to scaling and rotation. This prevents models from "cheating" by simply rescaling their weights to appear different while still learning the same underlying features. This feature-space repulsion allows models to find different "paths" to the correct reconstruction of normal data, ensuring they diverge when faced with out-of-distribution anomalies.
2. Dual-Space Uncertainty (DSU)
Standard anomaly detection methods typically only measure disagreement in the Output Space (the final reconstructed image). D2UE expands this by incorporating the Input Space. By calculating the gradient of the reconstruction error with respect to the input image, the system captures the model's internal "attention" or sensitivity. Even if two models produce similar reconstructions, their gradients may differ wildly on an anomaly. Combining these two spaces provides a much more robust signal for detection.
Experimental Results
The framework was evaluated across five diverse medical imaging benchmarks, including chest X-rays (RSNA, VinDr-CXR, CXAD), brain MRIs, and retinal fundus images (LAG).
Key findings include:
- State-of-the-Art Performance: D2UE consistently outperformed existing methods like DDAD and Multi-ST. It achieved an 88.6% AUC on the RSNA dataset and 86.3% AUC on the LAG dataset.
- Ablation Gains: On the RSNA benchmark, the combination of RAR and DSU improved the performance of a vanilla Autoencoder by 11.4% AUC, proving that both diversity and dual-space analysis are critical.
- Backbone Agnosticism: The method showed consistent improvements regardless of the underlying neural network architecture (e.g., AE, MemAE, or AEU).
Discussion and Future Work
The research highlights that Feature Space repulsion is more effective than Output Space repulsion. While the latter can lead to underfitting (where models fail to reconstruct even normal data), feature repulsion allows models to remain disciplined on normal data while diverging on anomalies.
Despite its success, the authors note that training an ensemble of $N$ learners increases computational overhead. Future work will likely focus on reducing the training cost and exploring the precise quantitative relationship between the degree of model diversity and the resulting performance gains in anomaly detection.
Summary updated