CLASSIFICATION OF TUMOR HISTOPATHOLOGY VIA SPARSE FEATURE LEARNING

Area of Science:

• Computational pathology
• Digital histopathology
• Machine learning in oncology

Background:

• Whole slide images (WSIs) require decomposition into distinct patches for linking histopathology to patient outcomes.
• Histology sections often originate from diverse laboratories with varying sample preparation protocols, posing analysis challenges.

Purpose of the Study:

• To develop and evaluate an unsupervised method for decomposing WSIs into distinct tissue components.
• To enable the linkage of histopathological features to clinical outcomes by analyzing large cohorts of histology sections.

Main Methods:

• Utilized a variation of the restricted Boltzmann machine (RBM) for unsupervised feature learning from image signatures.
• Employed computed codes from the learned representation to classify image patches.
• Evaluated the system on glioblastoma multiforme (GBM) and clear cell kidney carcinoma (KIRC) datasets from The Cancer Genome Atlas (TCGA).

Main Results:

• The RBM-based method successfully decomposed WSIs into clinically relevant regions (e.g., viable tumor, necrosis in GBM; tumor types, stroma in KIRC).
• Achieved high performance metrics: 84% for GBM and 81% for KIRC evaluations.
• Demonstrated the model's ability to characterize and visualize tumor architecture on large WSIs.

Conclusions:

• The unsupervised RBM approach provides an effective method for WSI decomposition and histopathological analysis.
• This technique facilitates the quantitative analysis of tumor architecture and its potential correlation with patient outcomes.
• The method shows promise for standardizing analysis across diverse histology datasets.