Adaptive Pseudo Label Selection for Individual Unlabeled Data by Positive and Unlabeled Learning

Takehiro Yamane, Itaru Tsuge, Susumu Saito, Ryoma BisearXiv 2025 · DOI · arXiv

Abstract

Semi-supervised learning (SSL) for medical image segmentation often relies on pseudo-labeling, which typically employs a uniform confidence threshold across all unlabeled data. However, medical images frequently exhibit diverse noise profiles and artifacts, making a global threshold suboptimal. This paper proposes an adaptive pseudo-label selection method using Positive and Unlabeled (PU) learning. By training image-specific classifiers on the feature distributions of individual unlabeled images, the method generates tailored metrics to distinguish foreground from background noise. Experiments on retinal vessel and 3D photoacoustic datasets demonstrate that this individualized approach significantly improves segmentation accuracy, approaching the performance of fully supervised models.

Introduction

Deep learning models for medical image segmentation, such as blood vessel identification, require extensive high-quality annotations, which are costly and time-consuming to obtain. Semi-supervised learning (SSL) addresses this by utilizing a small labeled dataset alongside a larger unlabeled pool. A prevalent SSL technique is pseudo-labeling, where a model trained on labeled data generates labels for unlabeled images.

A critical limitation of conventional pseudo-labeling is the use of a "one-size-fits-all" confidence metric. In medical imaging, scans vary significantly due to reconstruction artifacts, sensor interference, or patient motion. A global threshold suitable for clean images may be too lenient for noisy ones, introducing errors into the training process.

Comparison between conventional pseudo-labeling and the proposed adaptive method. Conventional methods use a fixed metric for all images, which may fail for specific noisy samples (Image A). The proposed method adapts metrics for each individual image.

Methodology

The proposed framework treats every unlabeled image as a unique case through a four-step pipeline:

  1. Pre-training: A standard segmentation network (e.g., U-Net) is trained on the available labeled data.
  2. Initial Pseudo-labeling: The pre-trained model generates confidence maps for unlabeled images. Pixels with very high confidence are assigned as initial foreground ($P_1$) and background ($N_1$) pseudo-labels.
  3. Adaptive PU Learning: For ambiguous pixels in a specific image, the system employs PU learning. It treats known foreground features from the labeled set as "Positive" and the ambiguous pixels as "Unlabeled." A small binary classifier is trained specifically on the feature distribution of that individual image, identifying reliable background labels ($N_{pu}$) tailored to its unique noise profile.
  4. Re-training: The final model is trained using the original labeled data combined with all selected pseudo-labels ($P_1, N_1, N_{pu}$).
The four-step pipeline of the proposed method: (1) Pre-train, (2) Initial pseudo-labeling, (3) Adaptive PU-Learning for image-specific negative labels, and (4) Re-training.

Experiments and Results

The method was evaluated on three datasets: the public DRIVE and CHASE DB retinal vessel datasets, and a clinical 3D Photoacoustic (PA) dataset for vascular structure estimation.

Key Results

  • Accuracy: On the DRIVE dataset, the proposed method achieved a Dice score of 0.796, outperforming the baseline (0.787) and standard pseudo-labeling (0.795), and nearing the fully supervised score of 0.805.
  • Clinical Utility: In 3D photoacoustic experiments, the "coverage" metric for vascular tracing improved from 0.712 (baseline) to 0.763.
  • Qualitative Performance: Visual analysis confirmed that the adaptive method effectively captures thin, low-contrast vessels often missed by conventional models.

Discussion

The core insight of this work is that individualization is crucial; applying PU learning to a batch of unlabeled data collectively yields inferior results compared to image-specific adaptation. This confirms that background noise in medical scans is highly idiosyncratic. PU learning is particularly effective here as it avoids the assumption that all non-positive pixels are negative, acknowledging an "unlabeled" category that prevents misclassifying small foreground structures as background.

Limitations and Future Work

Currently, the method focuses primarily on refining background (negative) pseudo-labels. Future research will explore adaptive metrics for foreground (positive) selection to identify "missed" pixels more accurately. Additionally, the authors plan to test the approach on other medical tasks, such as organ segmentation, and analyze its scalability with larger unlabeled datasets.

Summary updated

BibTeX
@misc{yamane2025adaptive,
  author = {Takehiro Yamane and Itaru Tsuge and Susumu Saito and Ryoma Bise},
  title = {{Adaptive Pseudo Label Selection for Individual Unlabeled Data by Positive and Unlabeled Learning}},
  year = {2025},
  eprint = {2508.07548},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.07548},
}
Figure