Source-linked AI summary

Label-Free Foundational Model Selection for Medical Image Classification under Distribution Shift via Pseudo Label Discrepancy

Juan Iñaki Larrea, Lucas Mansilla, Enzo Ferrante

arXiv:2608.25810v1cs.CV

TL;DR

Under inter-institutional distribution shift, choosing among foundation models is difficult because target labels are rarely available. The paper repurposes SUDO into AURCC, a label-free ranking criterion using unlabeled target predictions, and finds rankings that track target performance, especially when labeled source data are scarce.

  • Problem

    Under distribution shift, selecting the best foundation model for an unlabeled target domain is difficult because target-domain labels are rarely available.

  • Method

    AURCC partitions unlabeled target data by predicted probabilities, measures pseudo-label discrepancy within regions, and aggregates these discrepancies into a label-free model-ranking score.

  • Results

    Across three inter-hospital scenarios, AURCC strongly correlates with ground-truth target performance and is competitive with or superior to source-AUC ranking, particularly with small labeled sources.

  • Takeaways & Limitations

    AURCC reduces dependence on target labels for choosing among chest X-ray foundation models and is most useful when labeled source data are scarce.

  • Takeaways & Limitations

    Target subsampling to match the source positive rate was used as an experimental control, although it is not required by the method.

Abstract

from arXiv · show

Foundation models are increasingly deployed for medical image analysis. However, under the inter-institutional distribution shift typical of deployment, their performance varies widely and cannot be known without target-domain labels, which are rarely available. This leaves a practical question unresolved: given several candidate foundational models and labeled-data from a source domain, which one to deploy in an unlabeled target domain? We propose a label-free selection criterion built on SUDO, a framework for evaluating clinical AI systems without ground-truth annotations. SUDO partitions the unlabeled target data by predicted probability and, for each region, measures a pseudo-label discrepancy reflecting class contamination; aggregated across regions, this yields a score (AURCC) requiring neither target annotation nor fine-tuning. We show that AURCC can be used to rank a variety of vision-language models (BioMedCLIP, CXR-CLIP, CheXzero, MedCLIP, MedImageInsight, CLIP) on chest X-ray classification across three inter-hospital shift scenarios, under zero-shot and MLP-probe regimes. The AURCC ranking recovers the ground-truth ranking with Spearman rho up to 0.943 (p<0.05). Against the natural baseline of ranking by held-out source accuracy, AURCC is competitive when the labeled source is large and yields a more accurate ranking once it is small; the regime of interest in resource-constrained settings.

1 Introduction

Distribution shift can substantially degrade medical-image model performance, while target-domain annotations needed to measure that degradation are often unavailable. This work asks how to rank candidate foundation models for an unlabeled target domain using labeled source data.

  • Distribution shift arises when clinical images differ across scanners, hospitals, patient populations, or acquisition protocols, potentially degrading performance substantially.
  • Target-domain performance conventionally requires ground-truth annotations, but obtaining expert radiologist labels is costly and often inaccessible.
  • Vision-language foundation models support zero-shot and lightweight probe-based transfer with little or no supervision.
  • SUDO partitions unlabeled target data by predicted probabilities, measures pseudo-label discrepancy within regions, and aggregates these discrepancies into AURCC for model ranking.The proposed ranking uses source labels while target images only generate the probabilities driving partitioning.
  • The framework ranks candidate models by AURCC without target labels, using embeddings produced from labeled source and unlabeled target datasets.

2 Methodology

The method ranks pretrained foundation models under distribution shift by applying SUDO to target probabilities and aggregating regional pseudo-label discrepancies into AURCC. It evaluates zero-shot and MLP-probe probabilities across six chest X-ray models and inter-institutional datasets.

  • Each candidate foundation model receives an AURCC score computed without target labels, intended to approximate its true target-performance ranking.
  • SUDO ranking criterion: For each model, positive-class target probabilities are discretized into K quantile bins, each containing unlabeled target samples.
  • SUDO ranking criterion: Two auxiliary logistic regressions use opposing pseudo-label assumptions for each bin, producing held-out source AUC values s0(b) and s1(b).
  • SUDO ranking criterion: Large |SUDOb(ϕm)| indicates class dominance, whereas values near zero indicate contamination; scores are averaged over 300 stochastic runs.
  • SUDO ranking criterion: AURCC aggregates bin scores by tracing reliability against target coverage while sweeping symmetric quantile thresholds, yielding a score in [0, 1].
  • Models and regimes: The study compares zero-shot probabilities with a source-trained shallow MLP probe and evaluates BioMedCLIP, CheXzero, CXR-CLIP, MedCLIP, MedImageInsight, and CLIP-base.
  • Data and evaluation: The task is binary pneumonia classification across MIMIC-CXR, NIH ChestX-ray14, CheXpert, and PadChest, with PadChest serving as the shifted target domain.
  • Data and evaluation: Source data are split 80/20, supporting probe training and pseudo-label sampling on the training split and auxiliary evaluation on held-out data.

3 Results

AURCC closely recovers target-performance rankings across inter-hospital scenarios, while its advantage becomes clearer as held-out source labels shrink. The MP ranking is stable across region counts and outperforms source-AUC under the smallest tested sample sizes.

  • 3.1 Label-free model ranking: ρ = 0.943 (p < 0.05) for AURCC in every MP scenario and K value, indicating stable rankings across the number of probability regions.The comparison uses Spearman correlation with ground-truth AUCtarget.
  • 3.1 Label-free model ranking: AURCC orders models consistently with true target AUC, correctly placing the general-domain baseline below the medical models in all three scenarios.Intermediate-model ordering is not perfect: one pair is locally swapped in each scenario.
  • 3.2 Sensitivity to held-out source size: At N = 50, AURCC outperforms source-AUC in all three scenarios, including ρ = 0.683 versus ρ = 0.605 for MIMIC.For NIH at N = 100, AURCC achieves ρ = 0.911 versus ρ = 0.749 for source AUC.
  • 3.2 Sensitivity to held-out source size: AURCC consistently outperforms source-AUC for NIH and CheXpert at every tested held-out-source size, with wider gaps as the source shrinks.The sensitivity analysis averages Spearman ρ over 20 stratified subsampling repetitions.

4 Discussion

The study proposes AURCC as a label-free criterion for selecting foundation models under distribution shift, using unlabeled target images when source labels are scarce. Its demonstrated scope is chest-X-ray pneumonia classification with PadChest as the single target domain.

  • 4 Discussion: AURCC ranks foundation models using unlabeled target images and correlates strongly with ground-truth target performance across three inter-hospital scenarios toward PadChest.In MP, it is competitive with or superior to source-AUC, while in ZS the two criteria perform comparably.
  • 4 Discussion: AURCC outperforms source-AUC at every tested fraction for NIH and CheXpert and produces better rankings for N = 50 in all scenarios.The discussion attributes this signal to pseudo-label discrepancy across probability regions and target-distribution structure.
  • 4 Discussion: Generalization to other tasks, modalities, and multi-class settings remains to be validated because evaluation is restricted to pneumonia classification and PadChest.The three source datasets provide independent shift scenarios, but only one target domain is evaluated.
  • 4 Discussion: The pipeline uses embeddings and runs on CPU, enabling local model ranking on commodity hardware in resource-constrained settings.The paper frames this as reducing dependence on scarce target labels during model selection.
Loading 2608.25810v1…