Introduction
Digital pathology faces a significant "small $n$, large $p$" challenge. While biopsy slides contain millions of pixels, researchers often only have a single diagnosis per patient. This weak supervision makes training standard deep learning models difficult, as pixel-level annotations are rarely available. This paper proposes a Deep Multi-Instance Learning (MIL) framework to address these bottlenecks by treating slides as "bags" of small patches.
Methodology
The core of the proposed framework is the use of permutation-invariant operators. In the MIL paradigm, a slide is considered a "bag" and its constituent patches are "instances." A bag is labeled positive if at least one instance within it is positive.
To aggregate patch-level scores into a slide-level diagnosis, the authors evaluated three differentiable operators:
- Noisy-Or (NOR): A probabilistic logic operator where a bag is negative only if all patches are negative.
- Integrated Segmentation and Recognition (ISR): A ratio-based aggregator.
- Log-Sum-Exp (LSE): A smooth approximation of the maximum function.
Unlike traditional multi-stage MIL methods, this approach allows for end-to-end training via backpropagation, enabling the model to learn discriminative features directly from the data.
Datasets and Preprocessing
The study utilized 58 H&E stained histopathology images from 32 benign and 26 malignant breast cancer patients. Given the small dataset size, extensive data augmentation was employed to ensure model robustness.
As shown in the figure above, the preprocessing pipeline included H&E color decomposition to adjust stain intensity, random rotations, mirroring, and Gaussian blurring. These techniques help the model generalize better by simulating variations commonly found in clinical settings.
Results and Discussion
The Deep MIL models were compared against state-of-the-art Gaussian Process-based MIL (GPMIL).
| Method | Accuracy | AUC |
|---|---|---|
| DEEPNOR-MIL (Noisy-Or) | 0.879 | 0.88 |
| DEEPISR-MIL | 0.828 | 0.90 |
| RGPMIL (State-of-the-art) | N/A | 0.90 |
The results indicate that Deep MIL achieves performance comparable to state-of-the-art methods. Specifically, the Noisy-Or operator provided the most balanced performance in terms of accuracy and F-score. Furthermore, the model implicitly learns to identify Regions of Interest (ROIs) by assigning higher scores to patches that contribute most to the final diagnosis, effectively creating diagnostic heatmaps.
Conclusion
This work demonstrates that permutation-invariant operators in a Deep MIL framework can effectively classify histopathology slides under weak supervision. While the small dataset remains a limitation, the ability to train end-to-end and identify ROIs without pixel-level labels represents a significant step forward in automated cancer diagnosis. Future work may explore Bayesian deep learning to improve uncertainty quantification in clinical applications.
Summary updated