Computational textural mapping harmonises sampling variation and reveals multidimensional histopathological...

Area of Science:

• Histopathology
• Computational Biology
• Genomics

Background:

• Hematoxylin and eosin (H&E) digital slides face technical biases that compromise computational histopathology.
• Sample quality and variations introduce significant, undocumented technical fallacies in histopathology studies.

Purpose of the Study:

• To investigate if sample quality and sampling variation introduce technical fallacies in computational histopathology.
• To develop and validate a computational method for standardizing histopathological data and resolving technical biases.

Main Methods:

• Utilized The Cancer Genome Atlas (TCGA) clear-cell renal cell carcinoma (ccRCC) dataset, annotating ~78,000 image tiles.
• Trained deep learning models to detect histological textures and lymphocyte infiltration.
• Correlated histological features with clinical, immunological, genomic, and transcriptomic profiles.

Main Results:

• Deep learning models achieved 95% accuracy in classifying textures and lymphocyte infiltration.
• Texture analysis revealed sampling bias and suboptimal sample quality within the TCGA dataset.
• Computational texture mapping (CTM) normalized textural variance, harmonizing histopathological architecture.
• CTM-harmonized data revealed associations between tumor fibrosis, histological grade, EMT, mutation burden, and metastasis.

Conclusions:

• Texture-based standardization effectively resolves technical bias in computational histopathology.
• This approach facilitates understanding the molecular basis of tissue architecture.
• All developed code, data, and models are publicly released as a community resource.