OCTID

OCTID detects tumor-containing tiles in whole-slide hematoxylin and eosin (H&E) images using one-class learning to support patch-based cancer WSI analysis and is implemented as a Python package.

Key Features:

• Tile-level detection: Operates at the tile level to distinguish tumor-containing tiles from normal tiles in WSIs.
• One-class learning: Uses one-class learning to model normal tissue patterns from tumor-free WSIs and identify outlier (tumor) tiles.
• Pretrained CNN embeddings: Extracts features using a pretrained convolutional neural network (CNN).
• UMAP dimensionality reduction: Applies Uniform Manifold Approximation and Projection (UMAP) for feature space reduction.
• One-class SVM classification: Employs one-class support vector machines (SVMs) for final tile classification.
• Annotation-free operation: Defines non-tumor patterns using entire sets of normal WSIs, avoiding the need for detailed tumor annotations.
• Benchmark performance: Reported performance on four H&E image datasets with mean F1-score 0.90 ± 0.06, Matthews correlation coefficient 0.93 ± 0.05, and accuracy 0.94 ± 0.03.

Scientific Applications:

• Tumor tile selection: Identifies and excludes normal tiles to produce tumor-enriched training sets for patch-based models.
• Preprocessing for WSI analysis: Reduces dataset noise by removing normal tiles prior to downstream cancer WSI analyses.
• Annotation reduction in digital pathology: Minimizes reliance on slide-level or region-level tumor annotations by modeling normal tissue patterns.
• Benchmarking of tile detectors: Provides validated performance metrics for tile-level tumor detection on H&E datasets.

Methodology:

Features are extracted with a pretrained convolutional neural network, reduced using Uniform Manifold Approximation and Projection (UMAP), and classified by one-class support vector machines trained on tumor-free (normal) WSIs to identify and exclude normal tiles.