Deep Transfer Learning Across Cancer Registries for Information Extraction from Pathology Reports

Area of Science:

• Computational oncology
• Biomedical informatics
• Cancer surveillance

Background:

• Automated text information extraction from cancer pathology reports is crucial for national cancer surveillance.
• Developing robust information extraction tools that perform well across diverse cancer registries remains a significant challenge.

Purpose of the Study:

• To investigate the efficacy of transfer learning (TL) using convolutional neural networks (CNNs) for cross-registry knowledge sharing in cancer surveillance.
• To compare the performance of TL models with single-registry models and a cross-registry knowledge database.

Main Methods:

• Utilized transfer learning (TL) with a convolutional neural network (CNN) architecture.
• Trained and evaluated models using data from two distinct cancer registries.
• Focused on the information extraction task of primary tumor site and topography.

Main Results:

• Transfer learning (TL) demonstrated significant improvements in classification macro F-score compared to baseline single-registry models, with gains of 6.90% and 17.22%.
• The performance enhancement was particularly notable for classes with low prevalence, indicating improved handling of rare cancer types.
• Analysis confirmed TL's capability to facilitate knowledge sharing across registries.

Conclusions:

• Transfer learning (TL) with CNNs offers a viable strategy to enhance the generalizability and performance of information extraction tools for cancer surveillance.
• TL effectively addresses the challenge of cross-registry variability, leading to more robust models.
• The approach shows promise for improving the accuracy of cancer data collection, especially for underrepresented cancer types.