Abstract
Learning from noisy labels is a critical challenge in medical histopathology, where obtaining noise-free ground-truth labels is often impossible. This paper introduces One-Step Abductive Multi-Target Learning (OSAMTL), a framework that integrates machine learning with logical reasoning to handle complex label noise. By leveraging prior medical knowledge to abduce multiple plausible targets, OSAMTL constrains model predictions to be logically consistent. Furthermore, we propose the Logical Assessment Formula (LAF) to evaluate model performance based on logical rationality rather than unreliable noisy labels. Applied to Helicobacter pylori segmentation, OSAMTL demonstrates superior performance and logical consistency compared to state-of-the-art approaches.
Introduction
In medical histopathology whole slide image analysis (MHWSIA), experts often provide "noisy" labels—rough polygon annotations that may include background pixels or miss small details. This creates two major hurdles: traditional label correction methods fail against such complex, non-random noise, and standard evaluation metrics like IoU become unreliable without a clean gold standard. To bridge this gap, we propose a "Y-shaped" integration of perception (deep learning) and cognition (logical reasoning).
Methodology
One-Step Abductive Multi-Target Learning (OSAMTL)
OSAMTL moves beyond direct label correction by using Abduction—reasoning from observations to the most likely explanations. The process involves:
- Knowledge Base (KB) Integration: Incorporating medical facts (e.g., H. pylori are black, dot-like, and reside in luminal areas).
- Target Abduction: Generating multiple "abduced targets" from noisy labels and the KB. These targets represent different plausible interpretations of the truth, such as high-precision or high-recall variants.
- Multi-Target Learning: A joint loss function forces the deep learning model (e.g., a tiny FCN-based CNN) to remain consistent with all abduced targets simultaneously.
Logical Assessment Formula (LAF)
To solve the evaluation problem, the LAF framework measures Logical Rationality. Instead of comparing predictions to a noisy "ground truth," LAF calculates metrics like Logically True Positives (LTP) and Logically False Positives (LFP) by checking if the model’s output contradicts known medical facts. This provides a more robust measure of performance in real-world medical scenarios.
Experimental Results
The framework was evaluated using a dataset of 205 histopathology whole slide images from Peking University Third Hospital, processed into over 48,000 patches.
- Superior Performance: OSAMTL significantly outperformed state-of-the-art (SOTA) noisy-label methods. On the H. pylori task, it achieved a Logical F1-score (Lf1) of 67.56, compared to 42.72 for the baseline model trained directly on noisy labels.
- Rationality Gains: The most dramatic improvement was seen in the reduction of Logically False Positives (LFP), which dropped from 1550 in the baseline to 393 with OSAMTL. This indicates the model successfully learned to ignore "irrational" noise that contradicted medical logic.
- Compatibility: OSAMTL can be integrated into existing SOTA methods as a plug-in, consistently improving their logical consistency.
Conclusion
OSAMTL and LAF offer a novel path for training and evaluating models in domains where clean labels are unavailable. By treating logic as a constraint rather than just trying to "fix" labels, the framework achieves higher rationality in predictions. Future work will explore automated strategies for multi-target loss weighting and broader applications across different medical segmentation tasks.
Summary updated