Data Efficient and Weakly Supervised Computational Pathology on Whole Slide Images (2004.09666v2)

Abstract: The rapidly emerging field of computational pathology has the potential to enable objective diagnosis, therapeutic response prediction and identification of new morphological features of clinical relevance. However, deep learning-based computational pathology approaches either require manual annotation of gigapixel whole slide images (WSIs) in fully-supervised settings or thousands of WSIs with slide-level labels in a weakly-supervised setting. Moreover, whole slide level computational pathology methods also suffer from domain adaptation and interpretability issues. These challenges have prevented the broad adaptation of computational pathology for clinical and research purposes. Here we present CLAM - Clustering-constrained attention multiple instance learning, an easy-to-use, high-throughput, and interpretable WSI-level processing and learning method that only requires slide-level labels while being data efficient, adaptable and capable of handling multi-class subtyping problems. CLAM is a deep-learning-based weakly-supervised method that uses attention-based learning to automatically identify sub-regions of high diagnostic value in order to accurately classify the whole slide, while also utilizing instance-level clustering over the representative regions identified to constrain and refine the feature space. In three separate analyses, we demonstrate the data efficiency and adaptability of CLAM and its superior performance over standard weakly-supervised classification. We demonstrate that CLAM models are interpretable and can be used to identify well-known and new morphological features. We further show that models trained using CLAM are adaptable to independent test cohorts, cell phone microscopy images, and biopsies. CLAM is a general-purpose and adaptable method that can be used for a variety of different computational pathology tasks in both clinical and research settings.

• The paper’s main contribution is CLAM, a weakly supervised framework that uses attention-based pooling to aggregate patch features into effective slide-level representations.
• It achieves high AUC performance across tasks, reaching 0.991 for RCC, 0.956 for NSCLC, and 0.953 for lymph node metastasis detection.
• The model demonstrates robust adaptability and interpretability by generating diagnostic heatmaps that align with known morphological features used by pathologists.

Data Efficient and Weakly Supervised Computational Pathology on Whole Slide Images: An Expert Overview

In contemporary computational pathology, deep learning has emerged as a pivotal component enabling objective diagnosis and therapeutic predictions from whole slide images (WSIs). The paper presents CLAM (Clustering-constrained Attention Multiple instance learning), an advanced, high-throughput, and interpretable WSI-level processing methodology designed to mitigate the key challenges inherent in current deep learning approaches. The core contributions of this research lie in its weakly-supervised framework and data efficiency, alongside its adaptability and interpretability, heralding significant advancements in the field of computational pathology.

Methodology and Framework

CLAM leverages a deep-learning, weakly-supervised approach that obviates the need for pixel- or patch-level annotation. The framework employs attention-based pooling to aggregate patch-level features into integrated slide-level representations for classification. Specifically, CLAM executes attention-based learning to identify diagnostically significant regions within WSIs, assigning attention scores to patches that delineate their contribution to the slide-level diagnosis. This process is complemented by instance-level clustering to refine the feature space, which further bolsters the method's data efficiency and interpretability.

The model architecture of CLAM involves a dual-phase attention network wherein the first fully connected layer compresses patch-level representations into concise feature vectors. These vectors are then processed by parallel attention branches corresponding to each diagnostic class, culminating in class-specific slide-level representations. The final diagnostic prediction is derived by aggregating these representations using attention-pooling.

Performance Analysis

The authors systematically evaluate CLAM across three distinct pathology tasks: renal cell carcinoma (RCC) subtyping, non-small cell lung cancer (NSCLC) subtyping, and lymph node metastasis detection. The evaluation, conducted using both public datasets and independent test cohorts, reveals robust performance metrics. Notably, CLAM achieved a one-vs-rest mean test AUC of 0.991 for RCC subtyping, 0.956 for NSCLC subtyping, and 0.953 for lymph node metastasis detection at respective magnifications. The model demonstrated significant data efficiency, maintaining high AUC values even with reduced training set sizes, underscoring its utility in scenarios with limited data availability.

Generalization and Adaptability

One of the salient features of CLAM is its demonstrated adaptability. The performance remains consistent when applied to independent test cohorts, biopsy slides, and even cellphone-captured microscopy images. For instance, the models trained on resected tissue slides displayed high adaptability when tested on biopsy slides and cell phone microscopy images, achieving respectable AUCs without the necessity for further fine-tuning. This adaptability is crucial for real-world applications where data sources might vary significantly in imaging techniques and quality.

Interpretability

Another significant advantage of CLAM is its interpretability. The attention scores generated during the training process allow for the creation of heatmaps that visualize regions of high diagnostic importance within the tissue slides. These heatmaps align well with known morphological features used by pathologists, and in some instances, even delineate tumor boundaries without requiring explicit pixel-level annotation. This characteristic enhances the reliability of the model's predictions and provides a tool for pathologists to corroborate traditional diagnostic techniques with AI-driven insights.

Implications and Future Directions

The research presented in this paper holds substantial implications for both theoretical and practical aspects of computational pathology. The weakly-supervised approach, combined with data efficiency and adaptability, positions CLAM as a versatile tool capable of addressing a broad spectrum of pathology tasks. In practical terms, the method's ability to generalize across different data sources and conditions can lead to more widespread adoption in clinical settings. The research also opens pathways for further exploration into the discovery of novel morphological features and biomarkers, potentially augmenting diagnostic accuracies and treatment strategies.

Future work could focus on extending the CLAM framework to additional subtypes and pathologies, integrating domain-specific knowledge more deeply into the instance-level clustering process, and enhancing the interpretability features for even finer granularity in diagnostic insights. The open-source availability of CLAM on GitHub encourages community-driven advancements and adaptations, fostering collaborative progress in the field.

Conclusion

The paper introduces CLAM as a significant advancement in computational pathology, addressing critical challenges through an innovative weakly-supervised, data-efficient approach. The demonstrated performance, adaptability, and interpretability of CLAM affirm its potential to enhance diagnostic accuracy and efficiency in pathology, reaffirming the promising intersection of deep learning and medical imaging.