[Paper Review] Weakly-Supervised Deep Learning Model for Prostate Cancer Diagnosis and Gleason Grading of Histopathology Images

Mohammad Mahdi Behzadi, Mohammad Madani+7 more|arXiv (Cornell University)|Dec 25, 2022

AI in cancer detection4 citations

TL;DR

This paper proposes a weakly-supervised deep learning framework for prostate cancer Gleason grading using whole slide images (WSIs), leveraging a Transformer-based MIL model to identify discriminative patches, constructing a graph from high-scoring patches, and classifying grades via a gated-attention Graph Convolutional Network. The model achieves state-of-the-art performance on TCGA-PRAD, PANDA, and Gleason 2019 challenge datasets, with accuracy up to 0.912 and Cohen’s kappa of 0.892.

ABSTRACT

Prostate cancer is the most common cancer in men worldwide and the second leading cause of cancer death in the United States. One of the prognostic features in prostate cancer is the Gleason grading of histopathology images. The Gleason grade is assigned based on tumor architecture on Hematoxylin and Eosin (H&E) stained whole slide images (WSI) by the pathologists. This process is time-consuming and has known interobserver variability. In the past few years, deep learning algorithms have been used to analyze histopathology images, delivering promising results for grading prostate cancer. However, most of the algorithms rely on the fully annotated datasets which are expensive to generate. In this work, we proposed a novel weakly-supervised algorithm to classify prostate cancer grades. The proposed algorithm consists of three steps: (1) extracting discriminative areas in a histopathology image by employing the Multiple Instance Learning (MIL) algorithm based on Transformers, (2) representing the image by constructing a graph using the discriminative patches, and (3) classifying the image into its Gleason grades by developing a Graph Convolutional Neural Network (GCN) based on the gated attention mechanism. We evaluated our algorithm using publicly available datasets, including TCGAPRAD, PANDA, and Gleason 2019 challenge datasets. We also cross validated the algorithm on an independent dataset. Results show that the proposed model achieved state-of-the-art performance in the Gleason grading task in terms of accuracy, F1 score, and cohen-kappa. The code is available at https://github.com/NabaviLab/Prostate-Cancer.

Motivation & Objective

• To address the high cost and time burden of fully annotating whole slide images (WSIs) for prostate cancer diagnosis.
• To reduce interobserver variability in Gleason grading by automating the classification process.
• To improve classification accuracy by modeling spatial and relational dependencies among cancerous patches in WSIs.
• To develop a weakly-supervised approach that avoids patch-level annotations while maintaining high performance.
• To validate the model on multiple public datasets and an independent cross-validation set.

Proposed method

• Employing a Transformer-based Multiple Instance Learning (MIL) model to assign attention scores to image patches based on their contribution to the WSI-level label.
• Using an autoencoder to extract compact feature representations from patches, reducing computational load.
• Constructing a graph where nodes are high-scoring patches and edges connect each patch to its K nearest neighbors in spatial proximity.
• Applying a Gated-Attention Graph Neural Network (GNN) to classify the WSI into Gleason grades using node features and graph structure.
• Training the model end-to-end on bag-level labels (WSI labels), without requiring patch-level annotations.
• Using top-scoring patches from the MIL module as input to the graph construction, focusing on discriminative regions.

Experimental results

Research questions

• RQ1 Can a weakly-supervised deep learning model achieve high accuracy in Gleason grading without requiring patch-level annotations?
• RQ2 How does modeling spatial relationships between discriminative patches improve classification performance?
• RQ3 What is the optimal number of ASG modules and K-nearest neighbors for graph construction in this setting?
• RQ4 How does the percentage of selected high-score patches affect model robustness and accuracy?
• RQ5 Does the proposed framework outperform existing state-of-the-art models on public prostate cancer WSI datasets?

Key findings

• The proposed model achieved an accuracy of 0.912 and Cohen’s kappa of 0.892 on the Gleason 2019 challenge dataset, outperforming baseline models.
• On the PANDA dataset, the model reached 0.957 accuracy and 0.885 Cohen’s kappa, demonstrating strong generalization across datasets.
• The ablation study showed that selecting 60% of high-score patches significantly improved performance (from ~75% to ~90% accuracy) compared to 10%.
• Using 8 ASG modules yielded optimal performance across datasets, with 12 modules offering only marginal gains at higher computational cost.
• The KNN parameter had minimal impact on performance, with K=10 providing the best results across all datasets.
• The t-SNE visualization confirmed that the model learned discriminative, well-separated representations for different Gleason grades.