TCGA-Reports: A machine-readable pathology report resource for benchmarking text-based AI models.

Area of Science:

• Computational biology
• Medical informatics
• Oncology

Background:

• Pathology reports are rich, yet underutilized, data sources in cancer research.
• Existing structured data lacks the nuance and insights present in free-text pathology reports.
• There is a lack of publicly available, benchmark datasets for developing and evaluating pathology report analysis models.

Purpose of the Study:

• To create a machine-readable dataset of cancer pathology reports for AI-driven analysis.
• To establish a benchmark dataset for natural language processing (NLP) models in cancer research.
• To demonstrate the utility of the dataset for cancer-type classification.

Main Methods:

• Applied state-of-the-art optical character recognition (OCR) and custom post-processing to PDF pathology reports from The Cancer Genome Atlas.
• Generated a corpus of 9,523 machine-readable pathology reports.
• Performed proof-of-principle cancer-type classification across 32 distinct tissue types.

Main Results:

• Successfully created a comprehensive, machine-readable dataset of 9,523 pathology reports.
• Achieved a high performance of 0.992 average Area Under the Receiver Operating Characteristic curve (AU-ROC) in cancer-type classification.
• Demonstrated the feasibility of using NLP and AI for extracting valuable information from pathology reports.

Conclusions:

• The generated dataset addresses the need for benchmark resources in pathology report analysis.
• This dataset will facilitate advancements in AI and NLP applications for cancer research, clinical trials, and pathology.
• The findings highlight the potential of leveraging unstructured pathology data for improved cancer insights and diagnostics.