RPMIL: Rethinking Uncertainty-Aware Probabilistic Multiple Instance Learning for Whole Slide Pathology Diagnosis

Zhikang Zhao, Kaitao Chen, Jing ZhaoIJCAI 2025 · DOI

TL;DR

This paper introduces RPMIL, a novel probabilistic multiple instance learning (MIL) framework for whole slide image (WSI) pathology diagnosis. By modeling bag features as probability distributions rather than single point estimates, RPMIL captures data variability more effectively, leading to state-of-the-art performance on challenging WSI datasets.

What Problem Does This Paper Address?

Whole slide images (WSIs) are massive digital scans of pathology slides, crucial for cancer diagnosis. However, their gigapixel resolution and lack of fine-grained patch labels make traditional classification difficult. Multiple Instance Learning (MIL) is a common approach, where a WSI (bag) is divided into many patches (instances), and the presence of even one positive instance can label the entire bag as positive. Existing MIL methods typically extract instance features and aggregate them into a single, deterministic "bag feature" for prediction. This "point estimation" approach struggles to capture the full spectrum of data variability, especially when training data is limited, leading to fragile and potentially overfitted models. The paper aims to address this limitation by incorporating uncertainty estimation into the MIL framework.

Methods

The core of RPMIL is its probabilistic aggregator, which moves beyond deterministic bag features. Instead of mapping instances to a single bag representation, RPMIL learns a distribution of bag features. This is achieved using a variational autoencoder (VAE) that compresses multiple instance features into a low-dimensional latent space, representing the mean and variance of the bag feature distribution. This probabilistic representation allows for sampling multiple bag feature representations, providing a richer understanding of the data.

RPMIL also conceptualizes instance features within a bag as a distribution, jointly leveraging both instance and bag feature distributions for prediction. This approach is inspired by how pathologists analyze tumors by considering multiple areas. The model uses a cross-attention mechanism to fuse instance and bag features. To ensure the bag feature distribution remains stable and well-behaved, KL divergence and Mean Squared Error (MSE) losses are employed to regularize the distribution and prevent drift. The final prediction is made by integrating probabilities derived from both instance and bag feature distributions.

Datasets

The proposed RPMIL method was evaluated on two public datasets:

  • Camelyon16: A dataset for breast cancer classification, with 270 WSIs for training and 129 for testing.
  • TCGA-NSCLC: A dataset containing non-small cell lung cancer (NSCLC) WSIs, specifically lung adenocarcinoma (LUAD) and lung squamous carcinoma (LUSC). This dataset comprises 1053 WSIs, split into training, validation, and test sets in a 6:1.5:2.5 ratio.

Key Insights and Ideas

  • Uncertainty is Key: The paper strongly advocates for moving beyond point estimation in MIL. Modeling bag features as distributions, rather than single values, significantly improves performance by better capturing data variability and uncertainty.
  • Joint Distribution Power: Combining the distributions of both instance features and bag features for prediction yields superior results compared to relying on either distribution alone. This aligns with the intuition that considering individual tumor areas (instances) and their collective representation (bag) is crucial for accurate diagnosis.
  • Variance as a Signal: The variance in the bag feature distribution is shown to be a meaningful indicator of uncertainty. Removing samples with high variance improved performance, suggesting that the model can learn to down-weight or better interpret uncertain regions.
  • Sampling Matters: The number of samples drawn from the bag feature distribution during prediction has a direct impact on performance. A larger sampling size (e.g., 1,000) leads to more robust and accurate results, indicating that richer bag representations are obtained.

Key Results

RPMIL achieved state-of-the-art (SOTA) performance on both the Camelyon16 and TCGA-NSCLC datasets, outperforming various classical MIL methods.

  • On Camelyon16, RPMIL achieved an ACC of 91.72%, F1-score of 89.36%, and AUC of 95.13%. This represents a significant improvement over methods like DGMIL, with increases of 9.55% in ACC, 10.23% in F1-score, and 9.42% in AUC.
  • On TCGA-NSCLC, RPMIL achieved an ACC of 90.49%, F1-score of 89.34%, and AUC of 96.31%.
  • Ablation studies demonstrated that removing either the KL loss or MSE loss degraded performance, highlighting the importance of both regularization terms.
  • The paper also showed that the attention-based probabilistic aggregator combined with mean-pooling for bag features yielded the best results.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, probabilistic methods and sampling can inherently increase computational demands compared to deterministic approaches.
  • Hyperparameter Tuning: The performance of RPMIL relies on hyperparameters (λ₁, λ₂, λ₃) for the loss functions, which require careful tuning.
  • Future Work: The authors suggest that further exploration of uncertainty estimation in MIL, particularly in conjunction with advanced aggregator designs and potentially other probabilistic modeling techniques, holds promise for improving WSI pathology diagnosis. The visualization results also suggest that the model's attention mechanism is more focused on relevant pathological tissues compared to existing methods.

Summary updated

BibTeX
@inproceedings{zhao2025rpmil,
  author = {Zhikang Zhao and Kaitao Chen and Jing Zhao},
  title = {{RPMIL: Rethinking Uncertainty-Aware Probabilistic Multiple Instance Learning for Whole Slide Pathology Diagnosis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/275},
}
Figure