Risk-Aware Robust Learning: Reducing Clinical Risk under Label Noise in Medical Image Classification

Abstract: Noisy labels are a pervasive challenge in medical image classification, where annotation errors arise from inter-observer variability and diagnostic ambiguity. Although several noise-robust learning methods have been proposed, their evaluation predominantly relies on accuracy-oriented metrics, overlooking the clinical implications of asymmetric error costs. In medical diagnosis, a false negative (missed disease) carries substantially higher consequences than a false positive (false alarm), as delayed treatment can directly impact patient outcomes. In this work, we investigate whether noise-robust training methods preserve clinical safety under label noise. We conduct a systematic risk-aware evaluation of the state-of-the-art noise-robust methods Coteaching, DivideMix, UNICON, and a GMM-based filtering approach on binarized DermaMNIST and PathMNIST datasets under clean and label noise rates of 20%, and 40%. Beyond balanced accuracy, we adopt a cost-sensitive Global Risk formulation that explicitly penalizes false negatives. Our analysis reveals that the robustness of state-of-the-art methods does not guarantee clinical safety. Furthermore, we demonstrate that integrating cost-sensitive optimization into noise-robust training significantly reduces clinical risk, while mantaining model utility. These findings demonstrate that noise-robust learning must be evaluated through a clinical risk lens, and that combining robust training with cost-sensitive optimization can meaningfully reduce risk in noisy-label medical imaging scenarios.

• The paper proposes a cost-sensitive Global Risk metric that penalizes false negatives more heavily to better reflect clinical diagnostic safety.
• It rigorously evaluates robust training methods including UNICON+CS, Co-teaching, DivideMix, and GMM filtering under varying label noise conditions on DermaMNIST and PathMNIST.
• The results emphasize that high balanced accuracy may still entail unacceptable clinical risk, underscoring the need for risk-based evaluation in medical AI.

Risk-Aware Robust Learning under Label Noise in Medical Image Classification

Introduction

The challenge of label noise in medical image classification remains a significant reliability bottleneck for deploying deep learning systems in diagnostic applications. Annotation errors resulting from inter-observer variability and case ambiguity systematically introduce incorrect labels, which deep neural networks are prone to memorize, leading to serious misclassification risks. The cost asymmetry between false negatives (FN) and false positives (FP)—delayed disease detection versus unnecessary follow-ups—renders accuracy-oriented metrics insufficient for safety-critical deployment. This paper proposes a comprehensive clinical risk analysis of state-of-the-art (SOTA) noise-robust training methods, integrating a cost-sensitive Global Risk metric that penalizes false negatives commensurately, with particular focus on robust learning under varying label noise conditions.

Clinical Risk Formulation and Datasets

The authors employ DermaMNIST (skin lesion) and PathMNIST (colorectal pathology) datasets in their binary forms (malignant vs. benign), mirroring clinically meaningful screening scenarios. Label noise rates of 0%, 20%, and 40% are artificially introduced during training; test and validation splits remain clean. The evaluation eschews aggregate accuracy in favor of a risk metric defined as:

$$\text{Risk} = \frac{C_{FN} \cdot FN + C_{FP} \cdot FP}{N}$$

where $C_{FN}$ and $C_{FP}$ delineate the clinical costs—set to $C_{FN}=20$, $C_{FP}=1$ in their primary scenario (Risk II), thereby strongly penalizing missed malignant cases. This risk-based assessment is performed in parallel with classical metrics such as Balanced Accuracy (BAC), Sensitivity, and Specificity, to enable nuanced tradeoff analysis.

Approaches: Robust Training and Cost Sensitivity

Four robust learning algorithms—Co-teaching, DivideMix, UNICON, and a GMM-based filtering method—are evaluated against a cross-entropy baseline. Each is also paired with a cost-sensitive (CS) variant, implemented via class-weighted loss. Notably, UNICON augments uniform selection and contrastive learning, potentially providing more resilience for minority (malignant) class preservation under label noise. The cost sensitivity is designed to directly optimize for minimization of FN, correcting for the clinical risk asymmetry ignored by conventional loss formulations.

Empirical Evaluation: Clinical Risk under Noise

Systematic experiments quantify the interplay between label noise, robust training, and clinical risk. A pivotal observation is that increasing label noise destabilizes both BAC and risk metrics, but in dataset-dependent patterns. On DermaMNIST at 40% noise, the baseline model collapses to near-universal positive predictions, sharply reducing FN but inflating FP. This results in an artificial decrease of Risk II, while diagnostic utility is actually undermined—a degeneracy only visible when analyzing the error decomposition beyond aggregate risk.

Figure 1: Impact of label noise on the Baseline model; note the non-monotonic changes in Risk I and Risk II and the collapse in FN/FP balance at high noise.

Comparative results indicate that:

• UNICON+CS consistently achieves the lowest Risk II across both datasets and all noise rates, with an average Risk II reduction of 57% over the baseline (average of 1.25).
• Co-teaching exhibits the best BAC/Risk I performance at higher noise levels but demonstrates instability in sensitivity versus specificity as noise increases, particularly with PathMNIST at 40%.
• Methods like DivideMix and GMM Filter, in combination with CS loss, show mixed or degraded performance at higher noise, suggesting that naive integration of cost sensitivity with robust learning is non-trivial.

  Figure 2: Clinical Risk II comparison across methods, datasets, and noise levels. UNICON+CS attains the lowest risk in most settings.

Tradeoff Analysis: Accuracy vs. Asymmetric Risk

Tradeoff plots of BAC against clinical risk (Risk I and II) underscore that optimal accuracy does not guarantee clinical safety under high noise. In particular, models may exhibit high BAC but unacceptable Risk II (i.e., excessive FNs), directly contravening clinical objectives. The top-left region (high BAC, low Risk) is only reached by select method/cost combinations, reinforcing the argument for joint optimization.

Figure 3: BAC vs. Clinical Risk I. Optimal models reside in the upper-left quadrant.

Figure 4: BAC vs. Clinical Risk II. UNICON+CS situates closest to the optimal regime with low risk and sufficient BAC.

Methodological Implications and Theoretical Considerations

The findings reveal that robust learning and risk minimization are distinct and sometimes orthogonal: methods that maximize accuracy or resist noise do not necessarily minimize clinical risk due to asymmetric error costs. Effective risk reduction via cost-sensitive loss requires compatible robust learning mechanisms, notably those that ensure positive class retention under severe label corruption. Uniform class-aware sampling, as in UNICON, emerges as a key enabler.

A further implication is that risk-based metrics should be primary evaluation endpoints for medical AI development pipelines, with post hoc accuracy analyses providing secondary interpretability. Trivial strategies that minimize FN by predicting positives for all cases can result in misleadingly low clinical risk if specificity and further clinical workflow costs are ignored.

Prospects for Future Research

These results motivate several future directions:

• Development of theoretically principled, class-aware filtering strategies that integrate cost asymmetry in sample selection—beyond loss- or class-agnostic filtering.
• Adaptive tuning of cost ratios and loss weights to dataset-specific class imbalance and clinical context, potentially leveraging meta-learning.
• Extension of the framework to multi-class, hierarchical diagnostic settings, where cost matrices, rather than scalars, drive optimization.

Practical deployment in real-world medical systems mandates risk-aware protocol adoption for regulatory compliance and patient safety.

Conclusion

Noise-robust deep learning for medical image classification, if evaluated solely by symmetric metrics, is insufficient for the practical requirements of patient safety. This study robustly demonstrates that standard and SOTA robust training approaches may fail to optimize for clinical risk, particularly as label noise increases. Integrating cost-sensitive optimization with appropriate robust algorithms—specifically, those maintaining minority class representation—substantially improves real-world safety. Risk-aware optimization and evaluation are essential, and future noise-robust classification frameworks must internalize clinical asymmetries to ensure alignment with diagnostic utility.