Multiple Instance Learning with Mixed Supervision in Gleason Grading

Hao Bian, Zhuchen Shao, Yang Chen, Yifeng Wang, Haoqian Wang, Jian Zhang, Yongbing Zhang

Abstract

Gleason grading of prostate cancer using Whole Slide Images (WSIs) is a critical yet labor-intensive task for pathologists. While Multiple Instance Learning (MIL) has been widely adopted to handle the massive scale of WSIs, existing methods often struggle to effectively utilize both coarse slide-level labels and fine-grained, yet noisy, pixel-level annotations. This paper introduces a Mixed Supervision Transformer that leverages superpixel-based instance generation and a novel random masking strategy to bridge this gap. Our approach achieves state-of-the-art performance on the SICAPv2 dataset with a macro AUC of 0.9429, demonstrating significant robustness to annotation noise and the value of incorporating spatial context.

Introduction

The Gleason grading system is the gold standard for assessing prostate cancer severity. However, manual grading is time-consuming and prone to inter-observer variability due to the complexity of growth patterns across billions of pixels in a single Whole Slide Image (WSI). In machine learning, this is typically framed as a Multiple Instance Learning (MIL) problem, where a slide (bag) contains numerous tissue patches (instances). While slide-level labels are common, pixel-level annotations are scarce and often imprecise. Current MIL frameworks either rely solely on slide-level labels, missing local nuances, or over-fit to noisy pixel-level data. We propose a Mixed Supervision Transformer to address these challenges by effectively combining both label types.

Methodology

The proposed pipeline consists of two primary stages: instance generation and the mixed supervision transformer architecture.

  1. Superpixel-based Instance Generation: To better capture biological structures, we employ the Simple Linear Iterative Clustering (SLIC) algorithm to partition WSIs into superpixels rather than rigid rectangular patches. This ensures that instance-level labels align with natural tissue boundaries. Features for these superpixels are extracted using a MobileNetV2 backbone.
  2. Mixed Supervision Transformer: The core of our method is a dual-branch Transformer. The slide-level branch utilizes a class token to aggregate information for global grading, while the instance-level branch processes instance tokens to predict local Gleason grades.
  3. Random Masking and Spatial Encoding: To mitigate the impact of inaccurate manual annotations, we implement a random masking strategy, hiding a portion of instance tokens during training. Additionally, 2D sinusoidal position encodings are added to the tokens to provide the model with essential spatial context, mimicking a pathologist's consideration of tissue geography.
Overview of the Mixed Supervision Transformer pipeline, showing instance generation and the dual-branch architecture.

Experimental Results

We evaluated our model on the SICAPv2 dataset, which includes 155 slides with both slide-level Gleason scores and over 18,000 pixel-level annotated patches. Our method achieved a macro AUC of 0.9429, significantly outperforming existing slide-level supervision methods like CLAM and DSMIL, as well as advanced architectures like TransMIL.

Comparison of AUC performance against state-of-the-art MIL methods on the SICAPv2 dataset.

Discussion

A key insight from our experiments is the effectiveness of the random masking strategy. We found that masking 50% of instance labels actually improved performance compared to using the full set of annotations. This suggests that the model learns to generalize better by ignoring potentially noisy or overlapping human labels. Furthermore, the inclusion of 2D spatial encodings consistently boosted AUC, confirming that the spatial relationship between different tissue regions is vital for accurate Gleason grading. Qualitative analysis further validated that the model's predictions align closely with ground truth cancerous regions.

Conclusion

The Mixed Supervision Transformer offers a robust solution for prostate cancer grading by intelligently combining multi-scale labels. By respecting tissue morphology through superpixels and handling annotation noise via random masking, the model sets a new benchmark for the SICAPv2 dataset. Future work will focus on validating these results on larger multi-center datasets and enhancing the interpretability of the masking mechanism to provide clearer clinical insights.

Summary updated

BibTeX
@inproceedings{bian2022multiple,
  author = {Hao Bian and Zhuchen Shao and Yang Chen and Yifeng Wang and Haoqian Wang and Jian Zhang and Yongbing Zhang},
  title = {{Multiple Instance Learning with Mixed Supervision in Gleason Grading}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  doi = {10.1007/978-3-031-16452-1\_20},
}
Figure