[Paper] Self-Paced Learning for Images of Antinuclear Antibodies

Source: arXiv - 2511.21519v1

Overview

The paper introduces a self‑paced learning framework that can automatically detect antinuclear antibodies (ANA) directly from raw microscope images. By treating ANA detection as a multi‑instance, multi‑label (MIML) problem, the authors achieve state‑of‑the‑art performance without any hand‑crafted preprocessing, paving the way for faster, more reliable autoimmune‑disease diagnostics.

Key Contributions

• End‑to‑end MIML pipeline that works on untouched fluorescence microscopy images, eliminating costly manual preprocessing steps.
• Instance Sampler that filters out low‑confidence image patches by modeling pattern confidence, reducing noise from irrelevant regions.
• Probabilistic Pseudo‑Label Dispatcher that dynamically assigns soft labels to instances based on their visual distinguishability, mimicking how a human expert aggregates sub‑region observations.
• Self‑Paced Weight Learning that gradually adjusts instance importance during training, letting the model focus first on “easy” patterns before tackling harder, ambiguous ones.
• Comprehensive empirical validation on a proprietary ANA dataset and three public medical MIML benchmarks, achieving up to +7.0 % F1‑Macro and +12.6 % mAP improvements over the previous best methods.
• Open‑source implementation (GitHub) for reproducibility and rapid adoption.

Methodology

1. Problem Formulation – Each whole‑slide ANA image is treated as a bag of smaller patches (instances). The bag can contain multiple antibody patterns, so the task is inherently multi‑instance, multi‑label.
2. Instance Sampler – A lightweight confidence estimator scores each patch. Patches with low confidence are down‑weighted or discarded, preventing noisy background from contaminating the learning signal.
3. Pseudo‑Label Dispatcher – Instead of forcing a hard label on every patch, the dispatcher generates probabilistic pseudo‑labels that reflect how confidently a patch exhibits a particular ANA pattern. This mirrors a clinician’s “I see a hint of this pattern, but I’m not 100 % sure.”
4. Self‑Paced Learning (SPL) – Training proceeds in stages. Early epochs prioritize high‑confidence patches (the “easy” examples). As the model matures, the SPL scheduler gradually raises the weight of harder, ambiguous patches, allowing the network to refine its decision boundaries without being overwhelmed initially.
5. End‑to‑End Optimization – All three components are differentiable and integrated into a single deep‑learning backbone (e.g., ResNet). The whole system is trained jointly, so the sampler, dispatcher, and SPL coefficients co‑adapt to the data.

Results & Findings

The gains are consistent across diverse medical imaging domains, confirming that the self‑paced, pseudo‑labeling strategy generalizes beyond ANA detection. Ablation studies show that removing any of the three components (sampler, dispatcher, SPL) degrades performance by 4–9 %, underscoring their complementary roles.

Practical Implications

• Accelerated Diagnostics – Labs can replace manual slide‑by‑slide review with an automated system that delivers reliable ANA pattern reports in minutes, freeing up expert time for complex cases.
• Reduced Training Burden – Because the model learns directly from raw images, new labs can adopt the system without hiring staff to perform tedious image preprocessing or annotation standardization.
• Scalable to Other Multi‑Pattern Tests – The same framework can be repurposed for immunofluorescence assays (e.g., anti‑neutrophil cytoplasmic antibodies) or multiplexed pathology slides where multiple biomarkers coexist.
• Integration with Hospital Information Systems – The end‑to‑end nature allows the model to be wrapped as a micro‑service, feeding predictions straight into electronic health records (EHR) for real‑time decision support.
• Open‑Source Availability – Developers can plug the provided code into existing PyTorch pipelines, fine‑tune on institution‑specific data, or extend the sampler/dispatcher logic for new imaging modalities.

Limitations & Future Work

• Dataset Diversity – The primary ANA dataset originates from a single clinical center; broader multi‑center validation is needed to confirm robustness across different microscope brands, staining protocols, and patient populations.
• Label Granularity – While the pseudo‑label dispatcher handles ambiguity, the current system still relies on a fixed set of known ANA patterns; discovering novel or rare patterns would require additional unsupervised components.
• Computational Overhead – The instance sampling and SPL scheduling introduce extra forward passes per training batch, which may be a bottleneck for very large whole‑slide images on modest hardware.
• Future Directions – The authors suggest exploring self‑supervised pretraining on unlabeled microscopy data, extending the framework to 3‑D volumetric imaging, and incorporating active learning loops where uncertain patches are sent back to pathologists for targeted annotation.

If you’re a developer interested in trying out the code or adapting the pipeline to your own imaging workflow, the repository includes a ready‑to‑run Docker image and detailed instructions for training on custom datasets.

Authors

• Yiyang Jiang
• Guangwu Qian
• Jiaxin Wu
• Qi Huang
• Qing Li
• Yongkang Wu
• Xiao‑Yong Wei

Paper Information

• arXiv ID: 2511.21519v1
• Categories: cs.CV
• Published: November 26, 2025
• PDF: Download PDF