Comprehensive Word-Level Classification of Screening Mammography Reports Using a Neural Network Sequence Labeling...

Area of Science:

• Medical Imaging and Radiology
• Natural Language Processing
• Machine Learning

Background:

• Radiology reports contain valuable unstructured data.
• Current document-level classification methods lack granularity.
• Neural networks show promise in sequence-labeling tasks.

Purpose of the Study:

• To comprehensively classify mammography report text at the word level using a sequence labeling approach.
• To evaluate the performance of a neural network-based method against rule-based and CRF models.
• To demonstrate the utility of neural networks for granular data extraction from free-text radiology reports.

Main Methods:

• Developed a comprehensive 33-category classification system for screening mammography reports.
• Manually categorized words in 6705 reports, with initial pre-labeling by an algorithm.
• Employed a combined convolutional and recurrent neural network for word labeling and a siamese recurrent neural network for grouping findings.
• Compared performance against rule-based and conditional random field (CRF) models on an unseen test set.

Main Results:

• The neural network approach achieved significantly higher global accuracy (88.3%) compared to rule-based (57.0%) and CRF (75.8%) models (p < 0.001).
• Keyword accuracy for the neural network (95.5%) was also significantly higher than rule-based (80.9%) and CRF (76.9%) models (p < 0.001).
• Demonstrated superior performance in word-level multilabel classification across 33 classes.

Conclusions:

• Neural networks can accurately perform word-level multilabel classification of free-text radiology reports.
• A sequence labeling approach using neural networks is effective for granular NLP in radiology.
• This method significantly outperforms existing algorithms for comprehensive classification of screening mammography reports.