Deep Learning With Sampling in Colon Cancer Histology

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Genomic data analysis

Background:

• The Cancer Genome Atlas (TCGA) provides a rich resource for colon cancer research.
• Accurate cell identification is crucial for understanding tumor microenvironment and predicting patient outcomes.
• Traditional image analysis methods can be computationally intensive.

Purpose of the Study:

• To develop and apply a deep-learning algorithm for efficient cell identification in colon cancer diagnostic images.
• To investigate the association between cell-derived features and key clinical variables.
• To assess the performance and computational benefits of within-image sampling strategies.

Main Methods:

• A deep-learning cell identification algorithm was trained and applied to TCGA colon cancer images.
• Whole slide images were processed using tiling and cell identification within sampled patches.
• Two sampling policies, random and systematic random sampling, were evaluated for performance and accuracy.
• Derived cell features were statistically associated with clinical variables such as metastasis and invasion.

Main Results:

• Within-image sampling, particularly systematic random sampling, significantly reduced computation costs (seven-fold improvement) with minimal accuracy loss (~4%).
• Increased fibroblast counts correlated with metastasis, venous invasion, lymphatic invasion, and residual tumor.
• Decreased inflammatory cell counts were associated with mucinous carcinomas.
• Deep learning-derived features indicated cell density and cellularity, which are linked to patient survival.

Conclusions:

• Deep learning-based cell identification with efficient sampling is a powerful tool for colon cancer image analysis.
• Cellular features derived from this method show significant associations with critical clinical variables, offering prognostic potential.
• This approach enhances computational efficiency without compromising diagnostic accuracy, paving the way for broader clinical application.