NMGrad: Advancing Histopathological Bladder Cancer Grading with Weakly Supervised Deep Learning

TL;DR

The paper tackles inconsistencies in NMIBC grading and the scarcity of region-level annotations in WSIs. It proposes NMGrad, a pipeline that combines automatic urothelium segmentation, region-aware tile triplets at multiple magnifications, and a nested attention MIL (NMIA) for weakly supervised WHO04 grading. TRI-scale NMGrad achieves a high AUC of about 0.94 and provides interpretable region heatmaps, with an uncertainty spectrum further boosting F1 to 0.89 by filtering uncertain cases. The approach yields strong performance gains over prior methods and offers clinically relevant interpretability for bladder cancer grading.

Abstract

The most prevalent form of bladder cancer is urothelial carcinoma, characterized by a high recurrence rate and substantial lifetime treatment costs for patients. Grading is a prime factor for patient risk stratification, although it suffers from inconsistencies and variations among pathologists. Moreover, absence of annotations in medical imaging difficults training deep learning models. To address these challenges, we introduce a pipeline designed for bladder cancer grading using histological slides. First, it extracts urothelium tissue tiles at different magnification levels, employing a convolutional neural network for processing for feature extraction. Then, it engages in the slide-level prediction process. It employs a nested multiple instance learning approach with attention to predict the grade. To distinguish different levels of malignancy within specific regions of the slide, we include the origins of the tiles in our analysis. The attention scores at region level is shown to correlate with verified high-grade regions, giving some explainability to the model. Clinical evaluations demonstrate that our model consistently outperforms previous state-of-the-art methods.

Figures (12)

• Figure 1: NMGrad pipeline. Initially, we apply a tissue segmentation algorithm for ROI extraction. Then, we pinpoint diagnostically significant urothelium areas within WSIs. Subsequently, we split the urothelium mask into regions, based on proximity and size, and extract tile triplets. In a hierarchical fashion, we further transform these triplets within their corresponding regions into region feature embeddings using an attention-based aggregation method. All the region representations are then consolidated into a comprehensive WSI-level representation through a weight-independent attention module. Finally, this WSI feature embedding is input into the WHO04 grading classifier in order to produce accurate WSI grade predictions.
• Figure 2: We obtain comprising sets of three tiles at different magnification levels named triplets $\mathcal{T}$, enabling detailed examination. Tile triplets demonstrate regions associated with low- and high-grade features.
• Figure 3: Region definition. Urothelial tissue within a WSI is eligible for tile extraction. Blobs of tiles are formed, and blobs smaller than a threshold $T_\text{LOWER}$ are discarded. From the remaining blobs, any smaller than $T_\text{UPPER}$ are kept and defined as a region. For blobs bigger than $T_\text{UPPER}$, the blobs is subdivided into smaller pieces using the location of individual tiles within and KMeans clustering. The obtained clusters are designated as regions.
• Figure 4: Test performance for various aggregation techniques in weakly supervised learning. We provide the average of five runs, with the standard deviation shown in parentheses. The table presents the different approaches employed in the field, including aggregation techniques that involve considering spatial separation of the instances. We also explore the use of multiple magnification levels, considering 400x as the foundation for all magnification analysis. We also shown the results from other bladder cancer grading works, although in other datasets.
• Figure 5: Plot displaying the WSI predictions of the test set, with green shading representing the LG confidence interval, red for HG, and gray denoting the uncertainty interval. Additionally, a blue line depicts the regression line fitting the predictions.