Introduction
Training deep learning models for histopathological image analysis typically requires extensive manual annotation. In digital pathology, Whole Slide Images (WSIs) are gigapixel-scale files, making pixel-level manual tracing of tumor boundaries—a prerequisite for standard fully supervised segmentation—an incredibly time-consuming and non-scalable process. This paper addresses the "segmentation gap" by proposing a weakly supervised framework that leverages slide-level binary labels (tumor vs. normal) to train high-quality spatial segmentation models.
Methodology: The Proxy-Labeling Meta-Algorithm
The core innovation is a meta-algorithm designed to convert slide-level diagnostic labels into patch-level training signals using a Multiple Instance Learning (MIL) strategy.
The process follows four main steps:
- MIL Setup: A slide is treated as a "bag" of small image patches ("instances"). For a "normal" slide, all patches are assumed normal. For a "tumor" slide, at least one patch is tumorous.
- Inference: A ResNet50 backbone (pre-trained on ImageNet) performs a forward pass on a batch of patches.
- Proxy Labeling: Two hyperparameters, $\alpha$ and $\beta$, control the training signal. In a tumor-positive slide, the top $\alpha\%$ of patches with the highest predicted probabilities are assigned a "tumor" label (1), while the bottom $\beta\%$ are assigned a "normal" label (0). Patches in the middle are discarded to minimize noise.
- Optimization: Binary cross-entropy loss is computed only on these proxy-labeled patches, and the model is updated via backpropagation.
Experimental Setup and Datasets
The study utilized 6,481 flash-frozen WSIs from The Cancer Genome Atlas (TCGA), covering Kidney, Bronchus/Lung, and Breast tissue. Evaluation was conducted on:
- In-Distribution Test Set: 130 slides with expert manual annotations.
- Out-of-Location Test Set: 100 slides from unseen organs (Colon, Ovary, Uterus) to test generalization.
- Benchmark: The PatchCamelyon dataset was used to compare against fully supervised methods.
Key Insights and Results
The model's performance is highly sensitive to the control parameters $\alpha$ and $\beta$. The parameter $\alpha$, representing the minimum assumed tumor area, acts as a lever for recall; increasing $\alpha$ forces the model to identify more tumor tissue, often increasing recall at the cost of precision.
The best configuration ($\alpha=0.2, \beta=0.2$) achieved a pixel-wise AUC of 0.804 on the in-distribution test set. Remarkably, the model demonstrated "pan-cancer" generalization, maintaining an average AUC of 0.679 on unseen tissue types. On the PatchCamelyon benchmark, it reached an AUC of 0.802—a significant achievement for a model that never saw a pixel-level annotation during training.
Conclusion
This work demonstrates that "the power of extremes"—training on the most and least certain predictions—allows models to learn visual features of cancer through a self-correcting loop. While a performance gap remains compared to fully supervised models, this approach significantly lowers the barrier to entry for large-scale histopathological segmentation. The authors have released over 6,000 automatically generated tumor maps to support further research in the community.
Reference: Lerousseau, M., et al. "Weakly supervised multiple instance learning histopathological tumor segmentation." arXiv:2004.05024v4, 2021.
Summary updated