MARVEL: Margin-Aware Robust von Mises-Fischer Expert Learning for Long-Tailed Out-of-Distribution Detection

Technical Summary: MARVEL – Margin-Aware Robust von Mises-Fischer Expert Learning for Long-Tailed Out-of-Distribution Detection

Problem Statement
Deep neural networks (DNNs) have achieved high performance in medical imaging but face significant reliability challenges in real-world deployment. Two primary issues exacerbate this:

1. Out-of-Distribution (OOD) Detection Failure: DNNs often produce overconfident, erroneous predictions on OOD inputs (e.g., different anatomical regions, modalities, or non-radiographic images), posing safety risks.
2. Long-Tailed Class Imbalance: Medical datasets frequently follow a long-tailed distribution where a few "head" classes dominate and numerous "tail" classes are rare. This bias causes models to overfit to head classes, resulting in poor representation learning for tail classes. Consequently, tail samples are often misclassified as OOD, or true OOD samples are confused with tail classes due to overlapping feature clusters near decision boundaries.

Existing literature often assumes balanced datasets, evaluates OOD detection on coarse natural-image benchmarks, or fails to address the interplay between class imbalance and distributional uncertainty.

Methodology: MARVEL Framework
The authors propose MARVEL (Margin-Aware Robust von Mises-Fisher Expert Learning), a framework designed for medical imaging ID classification with OOD detection under long-tailed distributions. The architecture consists of a shared feature extractor and an ensemble of specialized experts.

1. Nonlinear von Mises-Fisher (NvMF) Classifier:

   • Motivation: Conventional vMF-based classifiers reduce to scaled cosine similarities, yielding linear decision boundaries in hyperspherical space. This limits flexibility for modeling complex, heterogeneous class geometries found in imbalanced data.
   • Innovation: The authors reinterpret the vMF distribution through its exponential-family structure. Instead of treating it solely as a likelihood model, they define logits using changes in the log-partition function ($\ell_c = -\Delta A(x)$).
   • Theoretical Guarantee: The paper proves (Theorem 1) that as the concentration parameter $\kappa \to \infty$, the proposed NvMF logits converge to standard cosine similarity, establishing it as a principled generalization of existing hyperspherical classifiers. For finite $\kappa$, it induces non-linear decision boundaries, allowing for better separation of complex clusters.

2. Margin-Aware Multi-Expert Ensemble:

   • To address long-tailed imbalance, the framework employs three NvMF experts, each trained with a distinct margin strength parameter $\tau \in \{0, 1, 2\}$:
     • Head-biased ($\tau=0$): Standard classifier for majority classes.
     • Balanced ($\tau=1$): Moderate rebalancing.
     • Tail-biased ($\tau=2$): Aggressive protection for minority classes.
   • Mechanism: The margin term $\Delta_{yc} = \tau \log(\pi_c / \pi_y)$ is introduced into the loss function, where $\pi$ represents empirical class priors. This asymmetrically shifts decision boundaries away from tail classes, preventing them from being overwhelmed by head classes.

3. Dedicated Outlier Expert:

   • A separate, simple fully connected layer ($g_{out}$) is trained explicitly for binary classification (ID vs. OOD).
   • It utilizes an auxiliary OOD dataset (ImageNet-100) during training. Unlike the NvMF experts which model fine-grained ID classes, this expert focuses solely on separating inliers from outliers, providing a direct supervision signal for OOD detection.

4. Inference Strategy:

   • ID Classification: Aggregates predictions from the three NvMF experts (averaging softmax probabilities over the $K$ ID classes).
   • OOD Detection: Combines two signals: (i) the probability of the $(K+1)$-th class (OOD) from the NvMF experts, and (ii) the explicit OOD probability from the outlier expert. The final score is the average of these two signals.

Evaluation Protocol
The authors critique existing benchmarks for relying on natural images (e.g., SVHN, LSUN) or simple class withholding. They propose a clinically realistic OOD evaluation framework spanning:

• Novel Classes: Semantic classes unseen during training (Open-Set).
• Corruptions: Synthetic degradations (noise, blur, artifacts).
• NearOOD: Domain shifts within medical imaging (e.g., different acquisition devices, staining techniques, or anatomical regions).
• FarOOD: Natural images (MS-COCO) visually distant from medical data.
• Datasets: Experiments are conducted on RFMiD (retinal imaging), ISIC2019 (dermatology), and NCTCRC (histopathology), all reshaped to follow a 50:1 long-tailed distribution.

Key Results

• OOD Detection: MARVEL consistently outperforms state-of-the-art methods (PASCL, COCL, EAT, PATT, OE) across all datasets and OOD scenarios.
  • It achieves mean FPR95 reductions of 8.45% (RFMiD), 13.02% (ISIC2019), and 36.90% (NCTCRC) compared to baselines.
  • It achieves near-saturated AUROC (>99%) on FarOOD and NearOOD2 scenarios, attributed to the dedicated outlier expert.
  • It maintains the best performance in the most challenging Open-Set scenario, though AUROC remains below 85% in this specific setting, highlighting the inherent difficulty of distinguishing novel classes from tail classes.
• ID Classification: MARVEL achieves the highest Accuracy and Balanced Accuracy. Unlike baselines that show a large gap between head and tail performance, MARVEL significantly improves tail-class recognition (e.g., tail accuracy approaches 95% on NCTCRC) while maintaining head-class performance.
• Ablation Studies:
  • Classifier Choice: NvMF significantly outperforms linear vMF, cosine, and FC classifiers, validating the benefit of non-linear boundaries.
  • Expert Count: An ensemble of three experts provides the optimal trade-off; adding a fourth expert yields no improvement and often degrades performance.
  • Calibration: MARVEL demonstrates superior confidence calibration compared to baselines, which tend to be either overconfident (PASCL, PATT) or underconfident (OE, EAT).
  • OOD Scoring: Logit-based scoring (MSP) within the framework outperforms distance-based methods (KNN, Mahalanobis), which struggle with the sparse, anisotropic nature of tail class representations.

Significance and Claims
The paper claims that MARVEL addresses a critical gap in medical AI by unifying long-tailed recognition and OOD detection. Its significance lies in:

1. Methodological Novelty: Introducing a theoretically grounded Nonlinear vMF classifier that generalizes cosine classifiers and a margin-aware multi-expert strategy to handle severe class imbalance.
2. Explicit OOD Modeling: Decoupling OOD detection from multi-class ID discrimination via a dedicated outlier expert, preventing the confusion between rare tail classes and true outliers.
3. Clinical Relevance: Moving beyond artificial benchmarks to evaluate models on a spectrum of clinically reflective distribution shifts (modality changes, acquisition variations, and natural images).
4. Safety: Enabling reliable identification of unfamiliar cases for deferral to clinicians, thereby supporting safer AI-assisted diagnosis in real-world workflows where data is inherently imbalanced and distribution shifts are common.

The authors acknowledge limitations, including the reliance on auxiliary OOD data for training and the assumption of a fixed label space, noting that future work could explore continual learning and dense prediction tasks.