Large-scale deep learning for metastasis detection in pathology reports.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Natural language processing for healthcare

Background:

• Existing algorithms struggle to reliably identify cancer metastasis from diverse pathology reports across the US population.
• Accurate identification of metastatic cancer is crucial for patient treatment and epidemiological studies.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated detection of metastatic cancer from unstructured pathology reports.
• To compare the performance of a task-specific deep neural network against a general-purpose large language model (LLM).

Main Methods:

• Utilized 60,471 unstructured pathology reports from Surveillance, Epidemiology, and End Results (SEER) registries.
• Developed a task-specific deep neural network trained from scratch.
• Coded reports into 'metastasis negative,' 'metastases positive,' or 'metastasis undetermined' categories.

Main Results:

• The deep learning model achieved a recall of 0.894, outperforming a general LLM's recall of 0.824.
• Quantifying model uncertainty allowed for deferring reports to human review, increasing recall to 0.969 when retaining 72.9% of reports.
• A smaller, task-specific deep learning architecture demonstrated superior performance over a general LLM.

Conclusions:

• Demonstrated the feasibility of using algorithms to automatically identify metastatic cancer cases from unstructured pathology reports.
• Highlighted the critical role of uncertainty quantification and abstention mechanisms in improving model performance.
• Suggests a pathway for more efficient and accurate cancer metastasis detection in clinical practice.