PaperlayerView exact revision on arXiv

Source-linked AI summary

Uncertainty of Vision Medical Foundation Models

Haoxu Huang, Narges Razavian

arXiv:2608.30390v1cs.LG

TL;DR

Medical AI needs uncertainty estimates that are reliable in high-stakes settings, but point probabilities can be mis-calibrated and lack formal coverage guarantees. This study evaluates point and region uncertainty across medical vision foundation models, pre-training sources, and calibration methods, finding that domain-specific self-supervised pre-training generally improves calibration and conformal efficiency while recalibration alone does not remove cross-model discrepancies.

  • Problem

    Point predictions can be mis-calibrated and provide no formal guarantees that prediction regions contain the ground truth in high-stakes medical applications.

  • Method

    The study evaluates medical vision foundation models using linear probing, point metrics, conformal region metrics, and recalibration techniques across pre-training sources.

  • Results

    Domain-specific self-supervised pre-training generally reduces point and region uncertainty, while standard recalibration does not fully resolve discrepancies across models trained on different data sources.

  • Takeaways & Limitations

    Reliable medical AI requires careful foundation-model selection together with both point- and region-level uncertainty quantification.

  • Takeaways & Limitations

    The evaluation focuses on disease classification and uses linear probing rather than full end-to-end fine-tuning.

Abstract

from arXiv · show

Accurate uncertainty estimation is essential for machine learning systems de- ployed in high-stakes domains such as medicine. Traditional approaches primarily rely on probability outputs from trained models (point predictions), which provide no formal guarantees on prediction coverage and often require additional calibra- tion techniques to improve reliability. In contrast, conformal prediction (region prediction) offers a principled alternative by generating prediction sets with finite- sample validity guarantees, ensuring that the ground truth is contained within the set at a specified confidence level. In this study, we explore the impact of pre-training approach, dataset scale and domain on both point and region-level uncertainty quantification, by studying domain-specific vision medical foundation models vs. general domain vision foundation models. We conduct a comprehensive evaluation across foundation models trained on retinal, histopathological, and Chest X-Rays data, applying various calibration techniques. Our results demonstrate that (1) pre-training on higher-quality domain-specific datasets along with self-supervised learning leads to better-calibrated point predictions than general domain pre-training, (2) stan- dard re-calibration methods alone cannot fully mitigate uncertainty discrepancies across models trained on different data sources, (3) domain-specific foundation model can lead to more efficient conformal prediction. These findings highlight the importance of careful model selection and the inte- gration of both point and region prediction to enhance the reliability and trust- worthiness of medical AI systems. Our work underscores the need for a holistic approach to uncertainty quantification in recent development of medical vision foundation model, ensuring robust and interpretable AI-driven decision-making.

1 INTRODUCTION

The section motivates reliable uncertainty estimation for high-stakes medical AI and introduces an evaluation of point and region predictions across foundation models and calibration settings.

  • Traditional probability outputs can be mis-calibrated and provide no formal guarantees that predictions cover the ground truth.This undermines the reliability of confidence estimates in high-stakes decision-making.
  • Conformal prediction generates prediction regions with finite-sample coverage guarantees without retraining the model.The region contains the ground-truth prediction with probability at least 1−α.
  • Prediction regions can represent possible classes, regression intervals, or pixel regions rather than a single point prediction.
  • Region-based AI systems combined with human decision-makers have shown potential to improve outcomes in randomized controlled trials.
  • The study investigates how diverse foundation-model data sources and calibration methods affect point and region uncertainty in medical vision models.The evaluation is framed around trustworthiness in critical decision-making.

2 RELATED WORKS

Prior work largely studies point-prediction uncertainty and methodological conformal prediction, leaving uncertainty across diverse foundation-model pre-training sources comparatively under-explored.

  • Softmax scores are commonly used as uncertainty proxies, but studies report that they are not well calibrated and motivate recalibration techniques.
  • Earlier uncertainty studies are constrained by limited data types, smaller-scale models, and emphasis on point rather than region prediction.
  • Prior findings disagree on whether model size or architecture families matter more for calibration and uncertainty.
  • Conformal-prediction research has addressed adaptive image classification, segmentation, and language-model problems mainly through algorithmic development.
  • Existing conformal-prediction work has limited exploration of models trained with different pre-training sources and methods.

3 PRELIMINARY

This section distinguishes point calibration from conformal prediction, which constructs coverage-guaranteed prediction sets and measures uncertainty through their size under exchangeability.

  • Calibration methods include post-hoc temperature scaling and regularization approaches such as entropy regularization.
  • Conformal prediction is a general framework for producing prediction sets with coverage guarantees rather than a specific algorithm.
  • A conformal score computed using held-out calibration data determines the threshold used to form C(X).
  • The prediction set C(X) maps an input to possible labels while targeting inclusion of the true label at confidence level 1−α.
  • The coverage guarantee relies on input exchangeability and does not require assumptions about the underlying model or data distribution.
  • Point uncertainty commonly uses ECE, Brier Score, and NLL, whereas conformal prediction directly measures uncertainty through prediction-set size.Smaller sets are described as more informative conformal predictors.

4 METHOD

The study evaluates medical vision foundation models primarily through linear probing, measuring both point- and region-level uncertainty and testing whether recalibration closes calibration gaps.

  • The study examines how domain-specific pre-training affects uncertainty estimation for vision foundation models in medical disease classification.The setup can extend to other use cases such as segmentation and report generation.
  • Linear probing trains only the classification layer to reduce overfitting-related overconfidence and accommodate models that expose generated features without weights.
  • Point uncertainty is evaluated with ECE, Brier Score, and NLL, while region uncertainty uses empirical coverage and set size.
  • The region-prediction evaluation applies Least Ambiguous set-valued Classifier and Regularized Adaptive Prediction Sets.
  • The pipeline tests whether recalibration techniques resolve calibration gaps after foundation-model pre-training.

5 DATA AND MODELS

The study evaluates seven ViT-Large foundation models across retinal, histopathology, and chest X-ray datasets, comparing domain-specific and general-domain pre-training sources and methods.

  • Models and evaluation: Seven ViT-Large foundation models are evaluated across retinal, histopathology, and chest X-ray imaging modalities.The models comprise four domain-specific and three general-domain foundation models.
  • Pre-training sources: The evaluation compares foundation models trained on different pre-training sources, including supervised, self-supervised, contrastive, and multimodal approaches.Examples include ImageNet21k, DINOv2, BioMedCLIP, RETFound, CTransPath, UNI, MRM, and Rad-DINO.
  • Domain-specific models: RETFound uses 1.6 million retinal images with self-supervised learning, while UNI uses 100 million histopathology images with distillation self-supervised learning.These models represent domain-specific pre-training at substantially different scales.
  • Domain-specific models: Rad-DINO is pretrained solely on large-scale chest X-ray data using DINOv2 image-only self-supervision.The model is designed without text supervision for biomedical image encoding.
  • Region prediction: Retina conformal prediction set size shows no clear trend across α thresholds by point-prediction uncertainty, pre-training source, or method.The figure reports average conformal prediction set size for retina data.

6 RESULT

The results show systematic differences in point-prediction uncertainty across pre-training choices, while calibration does not consistently remove those differences or predict conformal efficiency.

  • Point prediction uncertainty: Domain-specific pre-training consistently shows lower ECE before recalibration across datasets, with exceptions linked to possible pre-training–downstream distribution mismatch.Reported exceptions include UNI on TCGA and CTransPath on BraTS.
  • Point prediction uncertainty: Temperature scaling and label smoothing can reduce uncertainty in some cases but do not consistently close gaps between models with different initial uncertainty.Models with higher pre-calibration uncertainty generally remain more uncertain after recalibration.
  • Point prediction uncertainty: Supervised ImageNet21k models generally exhibit higher uncertainty than models trained with self-supervised approaches, even when the backbone remains unchanged during linear probing.This pattern persists across retinal, histopathology, and X-ray downstream tasks.
  • Region prediction uncertainty: Better point-prediction calibration after temperature scaling or label smoothing does not necessarily reduce conformal prediction set size.A well-calibrated model can still receive a similar (1−α) conformal quantile threshold.
  • Region prediction uncertainty: Histopathology and X-ray domain-specific models produce much smaller conformal prediction sets, whereas retinal domain-specific pre-training yields relatively large sets.UNI and Rad-DINO outperform other pre-trained models by a large margin, while RETFound offers mostly marginal improvement.

7 CONCLUSION

The paper concludes that medical foundation-model uncertainty should be evaluated at both point and region levels, combining model selection with uncertainty-aware methods.

  • Conclusion: Domain-specific pre-training combined with self-supervised learning generally reduces both point- and region-prediction uncertainty.The conclusion concerns medical image classification foundation models.
  • Conclusion: Temperature scaling and label smoothing improve calibration but do not fully resolve uncertainty differences among foundation models trained on different data sources.The conclusion supports evaluating calibration alongside foundation-model choice.
  • Conclusion: Conformal prediction efficiency cannot be directly inferred from point-prediction uncertainty calibration, despite its formal coverage guarantees.The paper presents conformal prediction as a robust framework for interpretable uncertainty estimation.
  • Conclusion: The paper advocates a holistic approach that combines careful foundation-model selection with rigorous uncertainty quantification for more reliable medical AI systems.The stated scope includes supporting safer and more informed clinical practice.

A DATASET

The datasets cover retinal diagnosis, histopathology subtyping, and chest X-ray classification, with multiple labeled collections within each modality.

  • Retina: The retina datasets include cataract and glaucoma-related labels, diabetic retinopathy stages, and blindness-assessment grades.Retina, IDRiD, and APTOS2019 provide four-, five-, and five-label tasks respectively.
  • Histopathology: CRC100K contains nine tissue subtypes from H&E-stained colorectal cancer and normal-tissue patches.The dataset uses non-overlapping histological image patches.

B METRICS AND ALGORITHMS

The section defines point-prediction calibration metrics and region-prediction methods for evaluating uncertainty. It also describes post-hoc calibration, conformal set construction, and empirical coverage assessment.

  • Point-prediction metrics: Expected Calibration Error (ECE) aggregates differences between accuracy and confidence across probability bins.ECE is sensitive to bin-size choices and loses information within bins.
  • Point-prediction metrics: Brier Score measures the mean squared difference between predicted probabilities and actual outcomes, capturing calibration and refinement together.Because it combines both properties, it is less suitable when they must be evaluated separately.
  • Point-prediction metrics: Negative Log-Likelihood evaluates probability assigned to the true class and penalizes incorrect or uncertain predictions more severely.It can produce misleading calibration assessments when the model is misspecified or probabilities are misaligned.
  • Calibration methods: Temperature scaling divides logits by a learned scalar from held-out validation data, while label smoothing mixes one-hot targets with a uniform distribution.Temperature scaling can miscalibrate under distribution mismatch, whereas label smoothing requires retraining and may cause under-confidence.
  • Region-prediction methods: LAC and RAPS construct conformal prediction sets, with RAPS regularizing set size and the experiments evaluating both set size and empirical coverage.The reported RAPS results coincide with the main LAC conclusion, while empirical coverage reaches at least 1 −α in all cases.
  • Region-prediction methods: LAC computes calibration scores, derives a quantile threshold, and includes labels whose scores do not exceed that threshold for a new example.This procedure controls the probability that the true label is excluded by α.

F POINT PREDICTION UNCERTAINTY RESULTS

The study reports point-prediction uncertainty using ECE, Brier Score, and NLL across retinal, histopathology, and Chest X-Ray datasets, with calibration variants included in the evaluation.

  • F POINT PREDICTION UNCERTAINTY RESULTS: The evaluation includes model outputs with linear probing, temperature scaling, and label smoothing.Figure 2 presents ECE results for these settings, while Brier Score and NLL results appear in the appendix.
  • F POINT PREDICTION UNCERTAINTY RESULTS: The datasets span retinal, histopathology, and Chest X-Ray domains.The corresponding dataset groups are listed in Tables 1 to 3.

G MODEL PERFORMANCE RESULTS

Model performance is reported across retinal, histopathology, and Chest X-Ray datasets using standard classification metrics. The calibration methods do not substantially change model accuracy performance.

  • G MODEL PERFORMANCE RESULTS: Calibration methods do not vary model accuracy performance much across the evaluated datasets.The comparison includes temperature scaling and label smoothing.
  • G MODEL PERFORMANCE RESULTS: Performance is evaluated with accuracy, balanced accuracy, AUROC, and AUPRC.Results are reported with means and 95% confidence intervals in Tables 4 to 6.
  • G MODEL PERFORMANCE RESULTS: The performance evaluation covers retinal, histopathology, and Chest X-Ray datasets.Tables 4 to 6 provide the corresponding domain-specific evaluations.
Loading 2608.30390v1…