Source-linked AI summary
Prevalence calibration as shortcut mitigation
Mohamed Amine Kina, Eike Petersen
TL;DR
Medical image classifiers can exploit shortcut features, while existing representation-focused mitigations have limited success and do not fit frozen encoders. This paper reframes shortcut learning as group calibration and proposes in-processing and post-hoc prevalence equalization; both substantially outperform baselines across experiments and model classes. On DenseNet, post-hoc recalibration raises misaligned-test AUROC from 0.23 to 0.73, though it slightly worsens (D)CSM.
Problem
Representation-level shortcut mitigation has limited empirical success and is incompatible with classifiers using frozen encoders.
Source: §1 · p. 2Method
The paper reframes shortcut learning as a calibration problem and proposes in-processing and post-processing methods that do not require modifying the encoder.
Source: §1 · p. 2Results
Both proposed methods substantially outperform all baselines across experiments and model classes; post-hoc recalibration raises standard ERM-trained DenseNet misaligned-test AUROC from 0.23 to 0.73.
Source: §5 · p. 7Takeaways & Limitations
The DenseNet result indicates that diagnostic features remain in the representation and that the classification head is the main bottleneck.
Source: §5 · p. 7Post-hoc recalibration substantially improves ERM but slightly worsens (D)CSM, a tension left for future work.
Source: §5 · p. 7
Abstract
from arXiv · showhide
Shortcut learning denotes the widespread situation in which a classifier exploits spurious correlations rather than diagnostic features. Existing mitigation strategies mostly aim to learn shortcut-invariant representations; their empirical success is limited and they cannot be applied to classifiers using frozen foundation model encoders. We propose to reframe shortcut learning as fundamentally a calibration problem: unconstrained learning implicitly calibrates each shortcut group to its training set disease prevalence, rendering the resulting classifier necessarily over-confident in one group and under-confident in the other. Building on this insight, we prevalence-equalize calibration between shortcut groups through two encoder-agnostic methods, an in-processing regularizer and a post-hoc prevalence-equalized recalibration step. Across chest-drain-pneumothorax benchmarks on CheXpert and SIIM-ACR, spanning fine-tuned CNNs and frozen foundation-model backbones, both methods substantially outperform all baselines. Post-hoc recalibration of a standard ERM-trained DenseNet raises misaligned-group AUROC from 0.23 to 0.73, indicating that shortcut reliance degrades the classification head rather than the underlying representation. Besides two new state-of-the-art shortcut mitigation approaches, our findings more fundamentally connect shortcut learning to calibration theory and algorithmic fairness.
1 Introduction
Medical classifiers can exploit spurious image features such as chest drains, and existing representation-level mitigations are incompatible with frozen encoders and may discard useful information. This paper reframes shortcut learning as group miscalibration and proposes encoder-agnostic mitigation methods.
1 Introduction
- Shortcut learning lets medical image classifiers exploit non-disease features correlated with labels, including chest drains in pneumothorax X-rays.Foundation-model encoders do not prevent classification heads from relying on such features.Source: §1 · pp. 1–2
- Representation-level mitigation is incompatible with frozen encoders and may discard shortcut-correlated information, potentially degrading downstream performance.The paper identifies these drawbacks as part of the motivation for an alternative mitigation approach.Source: §1 · p. 2
- Shortcut learning can be reframed as group miscalibration, motivating in-processing and post-processing methods that do not modify the encoder.The paper also reports extensive chest-drain pneumothorax experiments in which its methods outperform baselines by large margins.Source: §1 · p. 2
2 Background
Prior shortcut mitigation approaches use representation regularization, challenging-sample upweighting, last-layer retraining, or spatial specificity, with reported limitations. The background connects group calibration to prevalence and to fairness criteria that can be incompatible when group prevalences differ.
2 Background
- Existing shortcut mitigations include representation regularization, challenging-sample upweighting, retraining the final layer, and spatial specificity.Spatial specificity requires bounding-box annotations or segmentations; prior methods have achieved only partial success, with effectiveness varying across datasets and shortcut types.Source: §2 · p. 3
- Group calibration requires each group to be calibrated to its own prevalence, which explains its incompatibility with balance when prevalences differ.The passage also notes that calibration cannot hold under prevalence shifts without recalibration.Source: §2 · p. 3
- The background notes that implications of common group-fairness notions for shortcut mitigation have not been explored comprehensively.It situates this gap alongside work unifying causal formulations of shortcut learning and algorithmic fairness.Source: §2 · pp. 3–4
3 Methodology
The methodology builds shortcut-afflicted chest-drain pneumothorax datasets, then develops in-processing score matching and post-hoc group recalibration toward shared prevalence. It also specifies hyperparameter tuning and evaluation across CNN and frozen foundation-model settings.
3.1 Datasets
- CheXpert and SIIM-ACR splits use shortcut-afflicted training and validation data and test sets balanced across the four pneumothorax-by-drain combinations.CheXpert drain labels are inferred from a classifier because the dataset lacks native drain annotations; SIIM-ACR uses drain annotations and segmentation masks.Source: §3.1 · p. 4, Figure 2 · p. 5
3.2 Conditional score matching and prevalence-equalized calibration
- Class-Conditional Average Score Matching regularizes differences between groups’ average predicted scores within positive and negative disease strata.Its loss is the mean of the two strata’s across-group score standard deviations; dataset-level estimates provide a more robust alternative when batches lack all group-label combinations.Source: §3.2 · pp. 4–6
3.3 Post-hoc prevalence-equalized shortcut calibration
- Post-hoc recalibration fits separate shortcut-group mappings to a shared reference prevalence using beta calibration and inverse-class-probability validation weights.Shortcut-group membership must be available at inference time because the mappings differ by group.Source: §3.3 · p. 6
3.4 Hyperparameter optimization
- Hyperparameters are tuned to maximize the harmonic mean of pairwise validation AUROCs across positive and negative shortcut-group subsets.For a binary shortcut, this reduces to the harmonic mean of aligned- and misaligned-share AUROCs, penalizing gains in one group at the other’s expense.Source: §3.4 · p. 6
3.5 Experiments & evaluation
- Evaluation compares mitigation methods with ERM and other baselines on CheXpert DenseNet and frozen foundation models, and on SIIM-ACR DenseNet.Frozen foundation models are evaluated with methods operating only on the classification head; each experiment is run 10 times.Source: §3.5 · pp. 6–7
4 Results
The results show that in-processing methods and post-hoc recalibration improve shortcut mitigation benchmarks, with recalibrated ERM achieving the strongest reported outcomes.
4 Results
- 0.73: Post-hoc recalibration raises standard ERM from FS 0.37 to the overall-best result on CheXpert DenseNet.It exceeds DFR both before recalibration (0.56) and after recalibration (0.62).Source: §4 · p. 7
- DCSM and CSM lead the CheXpert DenseNet in-processing methods, substantially outperforming the next-best method, CDAN.Source: §4 · p. 7
- Frozen foundation models perform better overall, with further gains from DCSM and recalibrated ERM.Source: §4 · p. 7
- On SIIM-ACR, recalibrated ERM performs best, outperforming the bounding-box-dependent method of Saab et al..Source: §4 · p. 7
5 Discussion & Conclusion
The paper reframes shortcut learning as a calibration problem and reports that recalibration can recover performance on misaligned data. It also raises a group-calibration fairness implication while noting unresolved differences across methods.
5 Discussion & Conclusion
- ERM implicitly calibrates each shortcut group to its training prevalence, which differs strongly across groups.This reframes shortcut learning as a calibration problem and motivates in-processing and post-processing methods that substantially outperform baselines across experiments and model classes.Source: §5 · p. 7
- 0.23 to 0.73: Post-hoc prevalence-equalized recalibration raises DenseNet misaligned-test AUROC and aligned AUROC to 0.73, implicating the classification head as the main bottleneck.Diagnostic features remain in the representation; the recalibration result is consistent with DFR’s finding that last-layer retraining can recover robustness.Source: §5 · p. 7
- Group calibration to each group’s training prevalence requires different target priors; prevalence-equalized group calibration may be a more desirable fairness aim.The authors identify its implications as an area for future work.Source: §5 · p. 7
- Post-hoc recalibration substantially improves ERM but slightly worsens (D)CSM, a tension the authors leave for future work.Source: §5 · p. 7