TL;DR
By integrating Evidential Deep Learning into a self-supervised contrastive learning framework (SimCLRv2), this paper introduces an uncertainty-aware active learning strategy for digital pathology. The proposed method achieves state-of-the-art cancer subtyping performance using as little as 2-3% of expert-labeled data, drastically reducing the manual annotation bottleneck.
What Problem Does This Paper Address?
Deep learning has massive potential to streamline cancer diagnostics via digital pathology. However, training reliable models requires vast amounts of pixel-level annotations from expert pathologists—a process that is notoriously time-consuming, expensive, and prone to human variability.
While self-supervised learning (SSL) helps mitigate this by pre-training on unlabeled data, standard SSL models lack explainability and often struggle when applied to out-of-domain datasets. Active Learning (AL) introduces a human-in-the-loop approach to query only the most informative samples for labeling, but traditional AL querying strategies (like random sampling or basic entropy) often fail to select the best images or rely on computationally expensive uncertainty metrics like Monte Carlo dropout or deep ensembles.
This paper bridges the gap by embedding direct, mathematically rigorous uncertainty quantification directly into an advanced SSL pipeline to optimize active learning.
Methods
The authors propose UA-SimCLRv2 (Uncertainty-Aware SimCLRv2), a framework that combines self-supervised representation learning with Evidential Deep Learning (EDL).
- Self-Supervised Baseline: The architecture builds on SimCLRv2, which utilizes contrastive learning on unlabeled images, followed by supervised fine-tuning and knowledge distillation.
- Uncertainty Awareness: Instead of using a standard softmax layer for classification, the authors replace it with a ReLU activation layer to output a non-negative evidence vector. This vector parameterizes a Dirichlet distribution based on the Dempster-Shafer Theory of Evidence. This allows the network to model second-order probabilities and explicitly quantify epistemic uncertainty (the model's lack of knowledge) in a single forward pass.
- Active Learning Loop (UA-AL): The model starts by training on a tiny fraction (1%) of randomly labeled data. It then evaluates the remaining unlabeled pool, identifies the top 1% of images with the highest uncertainty scores, and requests expert labels for only those images. This process repeats iteratively up to a 10% labeling threshold.
Datasets
The framework was evaluated on two benchmark digital pathology datasets:
- Patch Camelyon (PCam): A binary classification dataset containing 327,680 histopathology patches extracted from lymph node sections, labeled for the presence or absence of metastatic tissue.
- NCT-CRC-HE-100K (NCT100k): A multi-class dataset consisting of 100,000 H&E-stained histological patches from human colorectal cancer and normal tissues, annotated into nine distinct tissue classes.
Key Insights and Ideas
- Uncertainty as a Proxy for Informative Data: The authors discovered a tight coupling between model errors and uncertainty. Plotting the histograms of uncertainty values revealed that correct predictions exhibit a left-skewed (low uncertainty) distribution, while incorrect predictions display a right-skewed (high uncertainty) distribution. This makes uncertainty an ideal metric for active learning queries.
- Cluster Border Refinement: t-SNE visualizations demonstrated that UA-SimCLRv2 refines cluster boundaries significantly better than standard SimCLRv2. The highly uncertain samples naturally lie at the ambiguous borders between tissue classes, meaning the active learning loop targets exactly the samples needed to resolve decision boundary confusion.
Key Results
- Label Efficiency: In multi-class colorectal cancer subtyping (NCT100k), the uncertainty-aware active learning framework achieved performance parity with state-of-the-art (SOTA) models using only 2-3% of the training labels.
- Surpassing the Baseline: When allowed just 9% of strategically selected labels, the out-domain UA-AL model achieved an accuracy of 97.01%, outperforming SOTA models (like HistoSSL-Res) trained on 100% of the dataset.
Limitations and Future Work
Despite its success, the authors note a limitation in the ultra-low label regime: at the absolute starting point of 1% labeled data, UA-SimCLRv2 occasionally underperformed compared to standard SimCLRv2. This is likely because the model lacks sufficient initial evidence to reliably quantify its own uncertainty.
Moving forward, the authors plan to explore the dynamics of in-domain versus out-domain pre-training more thoroughly. They also aim to scale this patch-level encoder up to Whole Slide Image (WSI) classification by integrating it as a backbone for Multiple Instance Learning (MIL) frameworks.
Summary updated