Use of BERT (Bidirectional Encoder Representations from Transformers)-Based Deep Learning Method for Extracting...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Natural Language Processing

Background:

• Liver cancer presents a significant health challenge in China.
• Dynamic contrast-enhanced computed tomography (DCE-CT) is crucial for liver cancer diagnosis.
• Radiology reports contain vital diagnostic information in free text.

Purpose of the Study:

• To develop an automated method for identifying liver cancer diagnostic evidence.
• To apply deep learning and rule-based natural language processing (NLP) to radiology reports.

Main Methods:

• A BERT-based BiLSTM-CRF model was developed to recognize hyperintense enhancement in the arterial phase (APHE) and hypointense in portal and delayed phases (PDPH).
• Traditional rule-based NLP methods were used to extract additional radiological features.
• A computer-aided diagnosis framework was designed using random forest with extracted features.

Main Results:

• The BERT-BiLSTM-CRF model achieved high F1 scores for APHE (98.40%) and PDPH (90.67%).
• A combined feature model yielded the highest performance (F1 score, 88.55%).
• APHE and PDPH were identified as the top two essential features for liver cancer diagnosis.

Conclusions:

• This study demonstrates a comprehensive NLP approach for liver cancer diagnosis from Chinese radiology reports.
• The BERT-based deep learning method achieved state-of-the-art performance in extracting diagnostic evidence.
• The findings support the extension of this deep learning method to other Chinese clinical texts.