Related Experiment Video
Updated: Dec 23, 2025
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
921
Comparison of deep learning models for natural language processing-based classification of non-English head CT
Yiftach Barash1,2, Gennadiy Guralnik3, Noam Tau1
1Division of Diagnostic Imaging, Sheba Medical Center, Sackler Faculty of Medicine, Tel Aviv University, Derech Sheba St 2, Ramat Gan, Israel.
Neuroradiology
|April 27, 2020
Summary
Deep learning models with word embedding significantly improve natural language processing for classifying non-English head CT reports. For specific tasks like hemorrhage detection, deep learning and traditional methods perform similarly.
Area of Science:
- Radiology and Medical Imaging
- Natural Language Processing
- Artificial Intelligence in Healthcare
Background:
- Natural Language Processing (NLP) offers potential for automating radiology report analysis.
- Classifying non-English radiology reports presents unique challenges for NLP models.
Purpose of the Study:
- To assess deep learning models for classifying non-English head CT reports.
- To compare the performance of different NLP models, including deep learning and traditional methods.
Main Methods:
- Retrospective collection of 176,988 Hebrew head CT reports (2011-2018).
- Manual labeling of 7784 reports for general (normal vs. pathological) and specific (intra-cranial hemorrhage) use cases.
- Implementation and evaluation of Long Short-Term Memory (LSTM) and LSTM-attention (LSTM-ATN) networks with Word2Vec embedding, compared against a Bag-of-Words (BOW) model.
Main Results:
- LSTM-ATN with Word2Vec achieved the highest AUC (0.967) and accuracy (90.8%) for general classification.
- For specific intra-cranial hemorrhage labeling, accuracies ranged from 95.0% to 95.9%, with LSTM-ATN-Word2Vec and BOW models showing the highest AUC (0.970).
Conclusions:
- Word embedding and attention mechanisms enhance NLP performance for general classification of non-English head CT reports.
- For specific tasks, traditional BOW models and deep learning approaches yield comparable results.
- Model selection and tailoring are crucial for optimizing NLP performance in radiology report analysis.
