PaperlayerView exact revision on arXiv

Source-linked AI summary

ACPL: Anti-curriculum Pseudo-labelling for Semi-supervised Medical Image Classification

Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, Gustavo Carneiro

arXiv:2111.12918v3cs.CV

TL;DR

Medical image SSL must work across multi-class and multi-label tasks despite class imbalance and pseudo-labelling’s associated challenges. ACPL selects informative unlabelled samples and uses an ensemble for pseudo labels, outperforming prior SSL methods on two benchmarks.

  • Problem

    Medical image SSL must accommodate multi-class and multi-label tasks with class imbalance, while pseudo-labelling can favor majority classes and reinforce incorrect predictions.

    Source: §1 · pp. 1–2

  • Method

    ACPL selects informative unlabelled samples for pseudo-labelling and combines deep-learning and KNN classifiers to estimate labels, with anchor set purification improving KNN pseudo-label accuracy.

    Source: §1 · p. 2

  • Results

    ACPL outperforms previous consistency-based, pseudo-label-based, and self-supervised methods on multi-label Chest X-ray14 and multi-class ISIC2018 benchmarks.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The experiments support ACPL as a pseudo-labelling approach for imbalanced medical image classification across multi-label and multi-class benchmarks.

    Source: §5 · p. 8

    The setup assumes labelled and unlabelled samples are drawn from the same latent distribution, with far fewer labelled samples.

    Source: §3 · p. 3

Abstract

from arXiv · show

Effective semi-supervised learning (SSL) in medical image analysis (MIA) must address two challenges: 1) work effectively on both multi-class (e.g., lesion classification) and multi-label (e.g., multiple-disease diagnosis) problems, and 2) handle imbalanced learning (because of the high variance in disease prevalence). One strategy to explore in SSL MIA is based on the pseudo labelling strategy, but it has a few shortcomings. Pseudo-labelling has in general lower accuracy than consistency learning, it is not specifically designed for both multi-class and multi-label problems, and it can be challenged by imbalanced learning. In this paper, unlike traditional methods that select confident pseudo label by threshold, we propose a new SSL algorithm, called anti-curriculum pseudo-labelling (ACPL), which introduces novel techniques to select informative unlabelled samples, improving training balance and allowing the model to work for both multi-label and multi-class problems, and to estimate pseudo labels by an accurate ensemble of classifiers (improving pseudo label accuracy). We run extensive experiments to evaluate ACPL on two public medical image classification benchmarks: Chest X-Ray14 for thorax disease multi-label classification and ISIC2018 for skin lesion multi-class classification. Our method outperforms previous SOTA SSL methods on both datasets

1. Introduction

Medical image SSL must accommodate multi-class and multi-label tasks with severe class imbalance. ACPL selects informative unlabelled samples and combines classifiers for pseudo-labelling, outperforming prior methods on two benchmarks.

  • Medical image SSL must handle both multi-class and multi-label tasks, which often have severe class imbalance from variable disease prevalence.Source: §1 · p. 1
  • Traditional pseudo-labelling can favor majority classes, is difficult to adapt across task types, and risks confirmation bias from incorrect predictions.Source: §1 · p. 2
  • ACPL selects informative unlabelled samples before pseudo-labelling, targeting likely minority-class examples without estimating class-wise thresholds.Source: §1 · p. 2
  • Informative mixup combines deep-learning and KNN classifiers to estimate pseudo labels.Source: §1 · p. 2
  • Anchor set purification adds informative pseudo-labelled samples to improve the KNN classifier’s pseudo-label accuracy.Source: §1 · p. 2
  • ACPL outperforms current state-of-the-art methods on Chest X-Ray14 multi-label and ISIC2018 multi-class classification.Source: §1 · p. 2

2. Related Work

Prior SSL work uses consistency learning, pseudo-labelling, and curriculum strategies, each with trade-offs for medical image tasks. ACPL’s informative-sample selection and classifier ensemble are presented as addressing these gaps.

  • Consistency-based SSL uses predictions across perturbed views, but requires domain-informed perturbation design that may need adaptation to each medical imaging type.Source: §2 · p. 2
  • Pseudo-labelling trains on high-confidence unlabelled predictions, but can bias imbalanced classification, complicate multi-class and multi-label use, and cause confirmation bias.Source: §2 · pp. 2–3
  • Anti-curriculum learning prioritizes hard or informative samples; ACPL applies it to pseudo-labelling to target likely minority classes without class-wise thresholds.Source: §2 · p. 3
  • ACPL surpasses consistency-based state-of-the-art methods on both chest X-ray and skin-lesion SSL benchmarks.Source: §2 · p. 3
  • ACPL also improves on computer-vision SSL state of the art applied to medical imaging, according to the paper.Source: §2 · p. 3

3. Methods

ACPL iteratively selects informative unlabelled samples for pseudo-labelling, combines model and KNN predictions, and purifies the anchor set. Its four modules use anchor-relative information content to guide selection, pseudo-labeling, and later training stages.

3.1. ACPL Optimisation

  • ACPL warm-starts on labelled data, then iteratively pseudo-labels selected unlabelled samples, trains with labelled and pseudo-labelled data, and updates the anchor set.The pseudo-labelled set starts empty and the anchor set starts with the labelled samples.Source: §3.1 · p. 4, Figure 2 · p. 4

3.2. Cross Distribution Sample Informativeness (CDSI)

  • CDSI estimates sample information content relative to anchor-set neighbours, selecting samples whose posterior probability of high information exceeds the low- and medium-category probabilities.The feature-density model is fitted with expectation-maximisation after anchor-set updates.Source: §3.2 · p. 4

3.3. Informative Mixup (IM)

  • Informative mixup forms pseudo-labels by combining model and anchor-set KNN predictions, with their contributions weighted by the sample's density score.The method addresses model confirmation bias while accounting for the KNN prediction's dependence on the initial labelled set.Source: §3.3 · pp. 4–5

3.4. Anchor Set Purification (ASP)

  • ASP limits anchor-set growth by inserting the least-connected pseudo-labelled samples, measured by how often they appear among neighbours of their anchor-set KNNs.The insertion threshold is the minimum connectivity count among pseudo-labelled samples.Source: §3.4 · p. 5, Figure 3 · p. 5

4. Experiments

ACPL is evaluated on multi-label Chest X-Ray14 and multi-class ISIC2018, with implementation details and ablations examining its training setup and proposed components. It outperforms comparison methods on both datasets, while high-information sample selection improves AUC and class balance in the Chest X-Ray14 ablations.

4. Experiments

  • The experiments use Chest X-Ray14 for multi-label classification and ISIC2018 for multi-class classification, with test metrics including AUC and, for ISIC2018, sensitivity and F1.Source: §4 · p. 5

4.1. Implementation Details

  • Both datasets use DenseNet-121; training warms up on labelled data before staged training with periodic anchor-set updates using ASP.For Chest X-Ray14, images are resized to 512 × 512; for ISIC2018, they are resized to 224 × 224.Source: §4.1 · pp. 5–6

4.2. Thorax Disease Classification Result

  • Across label percentages, ACPL leads on Chest X-Ray14 mean AUC, improving 2% over SRC-MT and NoTeacher and 1% AUC over S2MTS2 despite using DenseNet-121.Against UPS and GraphXNet, ACPL improves by 3% to 20%.Source: Table 1 · p. 6, §4.2 · pp. 6–7

4.3. Skin Lesion Classification Result

  • On ISIC2018, ACPL improves AUC by 0.7%–3% and F1 by around 1% over consistency-based approaches, and outperforms prior self-training across measures.Source: Table 3 · p. 7, §4.3 · p. 7

4.4. Ablation Study

  • At 2% labelled Chest X-Ray14, high-information selection yields around 73% AUC versus around 67% for low-information samples and 65% baseline, with a smaller labelled set and lower result variance.Source: §4.4 · p. 7, Table 4 · p. 7, Figure 4 · p. 7
  • ASP improves AUC by 0.3%–1.0% and reduces standard deviation by 0.4%–1.4%; with ASP, the best-to-worst AUC spread across K is around 1%, versus 2% without it.Informative Mixup produces pseudo-label results at least 1.5% better than the other tested approaches.Source: Figure 4 · p. 7, §4.4 · p. 8, Table 5 · p. 8
  • High-information selection shifts minority disease classes from 5%–10% to almost 30% and No Finding from 60% to 30%, balancing the five-class subset.Source: §4.4 · p. 8, Figure 5 · p. 8

5. Discussion and Conclusion

The conclusion presents ACPL as an SSL approach for imbalanced multi-class and multi-label medical image classification, reports results across two benchmarks, and identifies its in-distribution assumption as a boundary for future work.

  • ACPL combines informative unlabelled-sample selection with an ensemble of classifiers to produce pseudo-labels for imbalanced multi-class and multi-label MIA.Source: §5 · p. 8
  • ACPL outperforms prior consistency-based, pseudo-label-based, and self-supervised methods on multi-label Chest X-Ray14 and multi-class ISIC2018.Source: §5 · p. 8
  • The method assumes labelled and unlabelled samples are in-distribution; testing it with OOD data is future work.Source: §5 · p. 8

1. Additional Ablation study

The additional ablation compares the number of information-content sets on Chest X-Ray14 and weighs performance against training-stage count.

  • The default configuration divides unlabelled samples into three information-content sets: low, medium, and high.Source: §1 · p. 11
  • Using two sets gives the worst results because the pseudo-labelled set becomes less informative and imbalanced.Source: §1 · p. 11
  • Four sets perform similarly to three but greatly reduce new pseudo-labelled samples per stage and increase the number of stages, so three balances training time and accuracy.Source: §1 · p. 11

2. Data Distribution

The reported Chest X-Ray14 distributions show high-information selection producing a more balanced class mix than low-information selection or the original data.

  • High-information samples create a more balanced Chest X-Ray14 class distribution than low-information samples or the original data.Source: §2 · p. 11

3. Visualization of Classification Results

The section visualizes label distributions and density-mixup pseudo-label examples for Chest Xray-14 and ISIC2018. The figures distinguish information-content subsets and mark prediction accuracy.

3. Visualization of Classification Results

  • Figure 2 shows density-mixup pseudo-label examples for Chest Xray-14 and ISIC2018.The top and bottom panels correspond to Chest Xray-14 and ISIC2018, respectively.Source: §3 · p. 11, Figure 2 · p. 12
  • Figure 1 compares the original Chest X-ray14 label distribution with high- and low-information-content subsets across 14 classes plus No Finding.Green denotes the original distribution, blue the high-information subset, and yellow the low-information subset.Source: Figure 1 · p. 12
  • Green borders mark accurate predictions, red borders mark inaccurate predictions, and red class labels indicate ground truth in Figure 2.Source: Figure 2 · p. 12
Loading 2111.12918v3…