Comparison of deep learning models for natural language processing-based classification of non-English head CT reports.

Purpose: Natural language processing (NLP) can be used for automatic flagging of radiology reports. We assessed deep learning models for classifying non-English head CT reports.

Methods: We retrospectively collected head CT reports (2011-2018). Reports were signed in Hebrew. Emergency department (ED) reports of adult patients from January to February for each year (2013-2018) were manually labeled. All other reports were used to pre-train an embedding layer. We explored two use cases: (1) general labeling use case, in which reports were labeled as normal vs. pathological; (2) specific labeling use case, in which reports were labeled as with and without intra-cranial hemorrhage. We tested long short-term memory (LSTM) and LSTM-attention (LSTM-ATN) networks for classifying reports. We also evaluated the improvement of adding Word2Vec word embedding. Deep learning models were compared with a bag-of-words (BOW) model.

Results: We retrieved 176,988 head CT reports for pre-training. We manually labeled 7784 reports as normal (46.3%) or pathological (53.7%), and 7.1% with intra-cranial hemorrhage. For the general labeling, LSTM-ATN-Word2Vec showed the best results (AUC = 0.967 ± 0.006, accuracy 90.8% ± 0.01). For the specific labeling, all methods showed similar accuracies between 95.0 and 95.9%. Both LSTM-ATN-Word2Vec and BOW had the highest AUC (0.970).

Conclusion: For a general use case, word embedding using a large cohort of non-English head CT reports and ATN improves NLP performance. For a more specific task, BOW and deep learning showed similar results. Models should be explored and tailored to the NLP task.