Deep neural network models for computational histopathology: A survey (1912.12378v2)

Abstract: Histopathological images contain rich phenotypic information that can be used to monitor underlying mechanisms contributing to diseases progression and patient survival outcomes. Recently, deep learning has become the mainstream methodological choice for analyzing and interpreting cancer histology images. In this paper, we present a comprehensive review of state-of-the-art deep learning approaches that have been used in the context of histopathological image analysis. From the survey of over 130 papers, we review the fields progress based on the methodological aspect of different machine learning strategies such as supervised, weakly supervised, unsupervised, transfer learning and various other sub-variants of these methods. We also provide an overview of deep learning based survival models that are applicable for disease-specific prognosis tasks. Finally, we summarize several existing open datasets and highlight critical challenges and limitations with current deep learning approaches, along with possible avenues for future research.

• The paper presents a comprehensive survey reviewing over 130 deep neural network models applied to histopathological image analysis.
• It categorizes methodologies into supervised, weakly supervised, unsupervised, and transfer learning, highlighting advances in feature detection and image segmentation.
• The survey emphasizes the need for model interpretability and dataset diversity to improve clinical decision-making in digital pathology.

Deep Neural Network Models for Computational Histopathology: A Survey

This paper presents a thorough survey of deep learning techniques applied in the field of computational histopathology. It encompasses a review of over 130 papers and categorizes the literature based on methodological approaches, such as supervised learning, weakly supervised learning, unsupervised learning, and transfer learning. As these methods continue to evolve, the application of deep neural networks (DNNs) has become central to advancements in histological image analysis.

Supervised Learning

In supervised learning, the paper highlights the use of convolutional neural networks (CNNs) across a range of tasks such as classification, regression, and segmentation of histological images. These models excel in detecting specific histological features like nuclei and glandular structures. The survey notes the strong performance of multi-scale CNNs and models enhanced with transfer learning from pre-trained networks, significantly improving generalization across diverse datasets.

Weakly Supervised Learning

The survey describes weakly supervised learning as a pivotal technique to reduce the reliance on meticulously labeled data. Multiple-instance learning (MIL) methods are prevalent, particularly for weak label scenarios, allowing predictions at both instance and bag levels. This approach offers practical implications for clinical tasks, as it alleviates the exhaustive requirements for annotation by pathologists while maintaining robust detection capabilities.

Unsupervised Learning

Although in its nascent stages, unsupervised learning in histopathology is emerging as a promising field. Methods such as autoencoders and generative models have been used to discover intrinsic patterns in data without extensive labeling. This paper discusses the potential of these models to enhance cluster formations and latent feature extraction, which can improve the interpretability and efficiency of histological image analysis.

Transfer Learning and Domain Adaptation

Transfer learning is highlighted as a widely adopted practice in computational histopathology, leveraging pre-trained networks on large datasets like ImageNet. This method accelerates training and improves model convergence. Domain adaptation, particularly through adversarial networks and generative adversarial networks (GANs), is also discussed for handling domain shifts due to stain variability and other image acquisition discrepancies.

Survival Models for Prognosis

For prognostic tasks, the paper reviews deep learning models that incorporate survival analysis principles, such as Cox proportional hazards models integrated with CNNs. These models offer insights into predicting patient outcomes based on histopathological data, thus reflecting both theoretical significance and practical utility in real-world clinical decision-making.

Implications and Future Directions

The research emphasizes the importance of dataset diversity, model interpretability, and clinical relevance. As digital pathology continues to integrate into routine diagnostics, the development of comprehensive datasets and explainable models will be critical. The paper suggests that future research should focus on enhancing model transparency and reliability to gain clinician trust.

In summary, the survey provides an expansive view of current advancements in the application of deep learning for histopathology, underscoring both the successes and challenges of integrating these technologies into clinical settings. Moving forward, bridging the gap between academic research and practical implementation remains a key focus to fully realize the potential of AI in transforming pathology.