[Paper Review] An attention-based multi-resolution model for prostate whole slide imageclassification and localization

Jiayun Li, Wenyuan Li+4 more|arXiv (Cornell University)|May 30, 2019

AI in cancer detection39 references35 citations

TL;DR

The paper presents a two-stage attention-based MIL model for prostate whole slide image classification and weakly supervised ROI localization, achieving state-of-the-art accuracy on slide-level grading using only slide labels.

ABSTRACT

Histology review is often used as the `gold standard' for disease diagnosis. Computer aided diagnosis tools can potentially help improve current pathology workflows by reducing examination time and interobserver variability. Previous work in cancer grading has focused mainly on classifying pre-defined regions of interest (ROIs), or relied on large amounts of fine-grained labels. In this paper, we propose a two-stage attention-based multiple instance learning model for slide-level cancer grading and weakly-supervised ROI detection and demonstrate its use in prostate cancer. Compared with existing Gleason classification models, our model goes a step further by utilizing visualized saliency maps to select informative tiles for fine-grained grade classification. The model was primarily developed on a large-scale whole slide dataset consisting of 3,521 prostate biopsy slides with only slide-level labels from 718 patients. The model achieved state-of-the-art performance for prostate cancer grading with an accuracy of 85.11\% for classifying benign, low-grade (Gleason grade 3+3 or 3+4), and high-grade (Gleason grade 4+3 or higher) slides on an independent test set.

Motivation & Objective

• Address the challenge of prostate cancer grading on WSIs with only slide-level labels (weak supervision).
• Develop an MIL framework with attention mechanisms to identify informative tiles for grading and localization.
• Implement a two-stage multi-resolution approach to mimic pathologists' workflow (low-res screening then high-res grading).
• Demonstrate state-of-the-art accuracy on a large dataset and provide interpretable attention maps for ROI localization.

Proposed method

• Use a CNN backbone to extract instance-level features from tiles.
• Incorporate an attention module to compute instance weights and produce a bag-level embedding for slide classification.
• Apply instance dropout during training to encourage exploration of informative regions.
• Cluster instance features (PCA on reduced-dimension features) and aggregate attention by cluster to guide ROI selection.
• Operate a two-stage WSI pipeline: stage 1 classifies cancer vs non-cancer at 5x, stage 2 grades selected tiles at 10x into benign, low-grade, or high-grade.
• Transfer learning from a tile-level dataset to initialize feature extractors for both stages.

Experimental results

Research questions

• RQ1 Can weakly supervised MIL with attention localize informative regions in WSIs while achieving accurate slide-level grading?
• RQ2 Does a two-stage multi-resolution approach improve Gleason-grade classification compared to single-stage methods?
• RQ3 How does clustering-based ROI selection affect detection of informative regions and overall accuracy?

Key findings

• Achieved 85.11% accuracy on slide-level grading (benign, low-grade, high-grade) on an independent test set of 860 slides from 227 patients.
• Two-stage models with clustering-based attention outperformed one-stage and baseline attention models, with accuracy improvements over 7% relative to the one-stage baseline.
• Attention maps with instance dropout produced more comprehensive ROI localization than models without dropout.
• Compared to blue ratio-based ROI selection (br-two-stage), the attention-based ROI approach yielded higher accuracy (att-cluster-two-stage 85.11% vs br-two-stage 80.11%).
• Transfer learning from Cedars-Sinai tile dataset provided some benefits, but no-transfer model reached 84.30%.