Multi-scale Domain-adversarial Multiple-instance CNN for Cancer Subtype Classification with Unannotated Histopathological Images (2001.01599v2)

Abstract: We propose a new method for cancer subtype classification from histopathological images, which can automatically detect tumor-specific features in a given whole slide image (WSI). The cancer subtype should be classified by referring to a WSI, i.e., a large-sized image (typically 40,000x40,000 pixels) of an entire pathological tissue slide, which consists of cancer and non-cancer portions. One difficulty arises from the high cost associated with annotating tumor regions in WSIs. Furthermore, both global and local image features must be extracted from the WSI by changing the magnifications of the image. In addition, the image features should be stably detected against the differences of staining conditions among the hospitals/specimens. In this paper, we develop a new CNN-based cancer subtype classification method by effectively combining multiple-instance, domain adversarial, and multi-scale learning frameworks in order to overcome these practical difficulties. When the proposed method was applied to malignant lymphoma subtype classifications of 196 cases collected from multiple hospitals, the classification performance was significantly better than the standard CNN or other conventional methods, and the accuracy compared favorably with that of standard pathologists.

Cancer Subtype Classification Using Multi-Scale Deep Learning Approaches

The paper presents a novel convolutional neural network (CNN)-based methodology for classifying cancer subtypes from hematoxylin-and-eosin (H&E) stained histopathological images. The significant challenge addressed is handling whole slide images (WSIs), which are massive, typically surpassing dimensions like 40,000 x 40,000 pixels, including both cancerous and non-cancerous regions. Annotating tumor regions within these slides is not only laborious but also cost-prohibitive. Hence, the researchers devised a CNN architecture that integrates multiple instance learning (MIL), domain adversarial (DA) normalization, and multi-scale (MS) learning, aiming to mimic real-world pathological diagnostic practices.

Methodology

The proposed framework tackles several key difficulties inherent in histopathological image analysis:

1. Mixed Regions: WSIs typically contain a combination of tumor and non-tumor areas. The model identifies the regions containing tumor-specific features, leveraging MIL to focus on patches likely to contain relevant information without requiring explicit labels for each patch.
2. Variable Staining Conditions: Different stains from varied institutions can significantly impact image analysis. The use of DA normalization within the CNN helps mitigate this variability, allowing the model to learn features invariant to staining differences.
3. Scale Variation: Pathologists often change magnification levels to discern different tissue features. The novel multi-scale learning approach applies this practice algorithmically, analyzing WSIs at varying scales to uncover relevant diagnostic information.

The method was tested on 196 samples of malignant lymphoma, sourced from 80 hospitals, demonstrating substantial improvements over conventional CNN models and aligning closely with pathologists' diagnostic accuracy.

Experimental Results

The experimental setup included a binary classification task distinguishing between diffuse large B-cell lymphoma (DLBCL) and other lymphoma subtypes. The performance metrics indicated superior accuracy, precision, and recall for the proposed MS-DA-MIL method when compared with traditional patch-based CNN and single-scale MIL approaches.

Key numerical achievements included a top accuracy of 87.1% with multi-scale models, which outperformed single-scale models and conventional baselines. These results underscore the efficacy of integrating multiple learning paradigms to address diverse challenges in digital pathology.

Implications and Future Directions

The implications of this research are manifold. Practically, this approach could streamline the pathology workflow, reducing the need for labor-intensive manual annotations while maintaining diagnostic accuracy. Theoretically, the integration of domain adaptation, multi-scale, and multi-instance learning provides a robust framework for other complex pattern recognition tasks in biomedical imaging.

This work paves the way for further exploration into adaptive and unsupervised learning techniques that could enhance diagnostic models amid varying imaging conditions and heterogeneous datasets. Future research could expand on these models to support multi-class classification or even regression tasks like tumor grading, promoting more comprehensive AI-enabled diagnostic tools in clinical settings.

In conclusion, the integration of MIL with DA and MS within a CNN architecture represents a significant stride toward developing AI systems that can closely emulate human diagnostic strategies, potentially transforming histopathological analysis by bridging the gap between AI and clinical expertise.