Development and initial validation of a deep learning algorithm to quantify histological features in colorectal...

Area of Science:

• Computational pathology
• Digital pathology
• Oncology

Background:

• Colorectal carcinoma (CRC) diagnosis relies on histological assessment.
• Quantifying histological features in CRC is crucial for prognosis and treatment.
• Automated analysis of histopathology slides can improve diagnostic accuracy and efficiency.

Purpose of the Study:

• To develop and validate a deep learning algorithm for quantifying diverse histological features in colorectal carcinoma.
• To assess the algorithm's performance in segmenting tumor regions and classifying tumor characteristics.
• To evaluate the algorithm's ability to predict clinical outcomes such as metastasis.

Main Methods:

• A deep learning algorithm was trained on digitized hematoxylin and eosin-stained CRC slides (N=230).
• The algorithm segmented images into 13 regions and one object, achieving moderate to almost perfect agreement with pathologists.
• The validated algorithm was applied to an independent cohort (N=136) and analyzed features like tumor budding and lymphocytes.

Main Results:

• The algorithm accurately classified mucinous and high-grade tumors.
• Significant differences were identified between mismatch repair-proficient and mismatch repair-deficient (MMRD) tumors regarding mucin, stroma, and tumor-infiltrating lymphocytes (TILs).
• Algorithm-derived measures of tumor budding (TB) and poorly differentiated clusters (PDCs) outperformed routine grading and showed stronger associations with lymph node and distant metastasis.

Conclusions:

• Deep learning holds significant potential for identifying and quantifying a wide range of histological features in colorectal carcinoma.
• The developed algorithm demonstrates robust performance in classifying tumors and predicting metastatic potential.
• This technology could enhance the accuracy and efficiency of colorectal cancer diagnosis and prognostication.