Pathologist-level classification of histologic patterns on resected lung adenocarcinoma slides with deep neural networks (1901.11489v1)

Abstract: Classification of histologic patterns in lung adenocarcinoma is critical for determining tumor grade and treatment for patients. However, this task is often challenging due to the heterogeneous nature of lung adenocarcinoma and the subjective criteria for evaluation. In this study, we propose a deep learning model that automatically classifies the histologic patterns of lung adenocarcinoma on surgical resection slides. Our model uses a convolutional neural network to identify regions of neoplastic cells, then aggregates those classifications to infer predominant and minor histologic patterns for any given whole-slide image. We evaluated our model on an independent set of 143 whole-slide images. It achieved a kappa score of 0.525 and an agreement of 66.6% with three pathologists for classifying the predominant patterns, slightly higher than the inter-pathologist kappa score of 0.485 and agreement of 62.7% on this test set. All evaluation metrics for our model and the three pathologists were within 95% confidence intervals of agreement. If confirmed in clinical practice, our model can assist pathologists in improving classification of lung adenocarcinoma patterns by automatically pre-screening and highlighting cancerous regions prior to review. Our approach can be generalized to any whole-slide image classification task, and code is made publicly available at https://github.com/BMIRDS/deepslide.

• The paper presents a CNN-based model that achieves pathologist-level agreement with a kappa score of 0.525 and AUC values of at least 0.97.
• It employs a ResNet framework to segment whole-slide images into patches and aggregate predictions for predominant and minor histologic patterns.
• The model demonstrates potential to reduce diagnostic variability and augment pathologist workflows by pre-screening slides efficiently.

Pathologist-Level Classification of Histologic Patterns on Resected Lung Adenocarcinoma Slides Using Deep Neural Networks

The paper investigates the development of a deep learning model designed to classify histologic patterns of lung adenocarcinoma directly from whole-slide pathological images. The classification of these histological patterns is pivotal for determining the tumor grade and guiding treatment decisions for patients. Given the inherent variability and complexity in identifying these patterns, the potential applications of machine learning, particularly convolutional neural networks (CNNs), are substantial in reducing variability and increasing diagnostic consistency.

In terms of methodology, the paper uses a convolutional neural network, more specifically, a residual network (ResNet), to autonomously classify patches from lung adenocarcinoma slides. These individual predictions are aggregated to determine predominant and minor histological patterns for entire slides. The dataset tested includes slides from the Dartmouth-Hitchcock Medical Center, segmented into training, development, and testing datasets. The model's performance was evaluated against three pathologists, achieving a kappa score of 0.525 and agreement levels comparable to those of inter-pathologist evaluations.

The paper's results are notable in that the model operates at a level akin to experienced pathologists, achieving a kappa score and percentage agreement slightly superior to the pathologist inter-agreement. One of the significant numerical assessments indicated that the model accurately classified histologic patches with an AUC of at least 0.97 across all classes in a controlled development set, demonstrating the neural network's robust performance in distinguishing among histologic subtypes.

Key implications of this work lie in its practical utility— the model, if validated further, could supplement pathologists’ workflows by pre-screening slides, highlighting potential regions of interest, and potentially improving diagnostic turnaround times. Theoretically, this work contributes to the evidence supporting the use of AI in augmenting diagnostic procedures in pathology.

Future research directions suggested by the authors include testing the model in various clinical settings and incorporating its insights into the diagnostic pipeline. Moreover, extending this methodology to classify additional histological patterns or generalize to other whole-slide image tasks could provide further utility and validation for deep learning approaches in histopathology. Additionally, enhancing the model's granularity through techniques like region-based CNNs (R-CNN) could refine the localization of cancerous histological patterns and support more targeted pathology reviews.

Overall, this paper provides a comprehensive examination of leveraging deep learning for classifying lung adenocarcinoma patterns, demonstrating an alignment with, and potentially a supplement to, pathologists' expertise. This contributes to the ongoing dialogue and developments in applying machine learning to medical diagnostics, underscoring its potential role in enhancing accuracy and efficiency in clinical pathology.