OCTID: a one-class learning-based Python package for tumor image detection.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Patch-based analysis of whole slide images (WSIs) for cancer detection requires accurate tumor tile selection, a process that is currently labor-intensive and requires specialized expertise.
• While whole slides are often annotated as tumor or tumor-free, individual tiles within a tumor slide lack specific annotation, posing a challenge for automated analysis.

Purpose of the Study:

• To develop an automated method for accurate tumor tile classification in WSIs.
• To leverage tumor-free tiles for capturing normal tissue patterns using a one-class learning strategy, thereby reducing the need for manual annotation of tumorous regions.

Main Methods:

• Introduction of OCTID, a novel Python package for tumor tile classification.
• Integration of a pretrained convolutional neural network (CNN) model, Uniform Manifold Approximation and Projection (UMAP), and a one-class support vector machine (OCSVM).
• Training the model using a dataset exclusively composed of tumor-free tiles to learn normal tissue characteristics.

Main Results:

• The OCTID package demonstrated high performance in tumor tile classification across four H&E image datasets.
• Achieved remarkable performance metrics including an F1-score of 0.90 ± 0.06, a Matthews correlation coefficient (MCC) of 0.93 ± 0.05, and an accuracy of 0.94 ± 0.03.
• The one-class learning approach effectively utilized tumor-free tiles for classification.

Conclusions:

• OCTID provides an efficient and accurate solution for automated tumor tile classification in digital pathology.
• The developed method reduces the reliance on manual expertise and labor for WSI analysis.
• The package offers a valuable tool for advancing computational pathology research and applications.