PathCNN

PathCNN classifies pan-cancer whole-slide images (WSIs) using a simplified convolutional neural network to enable automated tumor-type and outlier detection across multiple tumor sites and non-neoplastic tissue.

Key Features:

• Simplified CNN Architecture: A streamlined convolutional neural network architecture contrasted with more complex models such as Google's Inception to reduce model complexity while maintaining or improving classification performance.
• Pan-Cancer Classification: Trained on whole-slide images from multiple tumor sites and corresponding non-neoplastic tissue to classify a wide range of cancer types and subtypes.
• Dimensionality Reduction of Last-Layer Weights: Analysis of the network's last-layer weights groups images by cancer type, highlights staining technique variations, and aids identification of outliers, artifacts, and potential misclassifications.
• Outlier and Artifact Detection: Detects anomalies such as artifacts and image-processing errors within WSI datasets to flag atypical or low-quality samples.
• Computational Efficiency: Optimized for more efficient computational resource utilization compared with more complex architectures while preserving classification accuracy.

Scientific Applications:

• Oncological research: Enables pan-cancer studies by classifying WSIs across tumor sites and non-neoplastic tissue for comparative and large-cohort analyses.
• Clinical diagnostics: Supports tumor-type and subtype classification on whole-slide images and assists in identifying staining-related issues and potential misclassification relevant to diagnosis.
• Quality control and dataset curation: Automates detection of artifacts, outliers, and staining variability in large WSI datasets to improve data quality for downstream analyses.

Methodology:

PathCNN trains a simplified convolutional neural network on extensive whole-slide image datasets from multiple tumor sites and non-neoplastic tissues and applies dimensionality reduction analyses to the network's last-layer weights to visualize groupings and detect outliers.