Study on structured method of Chinese MRI report of nasopharyngeal carcinoma

Xin Huang, Hui Chen, Jing-Dong YanBMC Medical Informatics Decis. Mak 2021 · DOI

Abstract

Medical imaging reports for nasopharyngeal carcinoma (NPC) are predominantly stored as unstructured free-text, hindering automated analysis and cross-institutional data sharing. This study proposes an end-to-end deep learning pipeline to structure Chinese MRI reports. By utilizing a character-level BERT-CRF model, the approach avoids the common pitfalls of traditional Chinese word segmentation. A novel Primary-Subsidiary-Attribute-Value (PSAV) triplet architecture is introduced to capture complex anatomical relationships. Results demonstrate that the BERT-CRF model achieves 99% accuracy in entity recognition, enabling the construction of a comprehensive medical knowledge network for clinical decision support.

Introduction

While medical records are increasingly digitized, radiology reports often remain as free-text narratives. This unstructured format creates "information islands," making it difficult for computers to parse or analyze data at scale. Structuring Chinese medical text is particularly challenging because the language lacks spaces between words. Traditional Natural Language Processing (NLP) methods rely on word segmentation tools that often fail when faced with the abbreviations and non-standardized terminology common in clinical settings. A single error in segmentation can cascade, ruining downstream medical entity extraction.

Methodology

The researchers developed a pipeline that bypasses word segmentation entirely, opting for character-level Named Entity Recognition (NER) combined with a custom rule base.

Figure 1: Overall framework of the proposed method, illustrating the stages of data preprocessing, model training, and network construction.

Expanded Triplet Architecture

Standard <entity, attribute, value> triplets are often insufficient for the hierarchical nature of radiology. The authors expanded this to a four-part triplet: <(Primary entity, Subsidiary entity), Attribute, Value> (PSAV). For example, a report might be structured as: Primary = "Nasopharyngeal cavity", Subsidiary = "Tumor", Attribute = "Size", Value = "1.2 cm x 0.9 cm".

NER and Rule-Based Structuring

The extraction was framed as a sequence labeling task using the BISO (Begin, Inside, Single, Outside) tagging scheme. The study compared five neural network architectures: CNN-CRF, IDCNN-CRF, LSTM-CRF, BiLSTM-CRF, and BERT-CRF. To handle compound sentences (e.g., "no abnormalities in ethmoid and maxillary sinus"), a conjunction rule database was implemented to ensure values were correctly mapped to all distinct entities.

Datasets

The study utilized a proprietary dataset of 769 desensitized MRI reports of nasopharyngeal carcinoma from a large tertiary hospital in China. The data was split into a 6:2:2 ratio for training (461 samples), validation (154 samples), and testing (154 samples).

Key Results and Insights

  • Superiority of BERT-CRF: The BERT-CRF model outperformed all other variants, achieving 99% accuracy and recall at the entity recognition level. This confirms that character-level deep learning effectively eliminates the need for error-prone word segmentation.
  • Pipeline Performance: When applied to full sentences to extract complete PSAV triplets, the overall pipeline achieved a structured information extraction rate of 84.74% and a structure accuracy of 89.39%.
  • Hybrid AI Approach: While deep learning handled the high-variance task of entity recognition, a deterministic rule base proved highly effective for resolving the logic of conjunctions, demonstrating the value of combining neural networks with symbolic logic.
  • Knowledge Graph Generation: The system successfully generated a visual knowledge network mapping primary entities to their observed states, providing an intuitive, structured view of NPC clinical characteristics.

Discussion and Future Work

Despite the high accuracy, the BERT-CRF model is computationally intensive, which may pose challenges for deployment in resource-constrained environments. Additionally, the algorithm still faces difficulties with highly complex compound sentences describing multiple distinct findings (e.g., multiple lymph nodes of varying sizes).

Future research will focus on expanding the dataset to include multi-institutional data to test generalizability, improving the handling of complex sentence structures, and normalizing synonymous terms to further refine the medical knowledge network.

Summary updated

BibTeX
@article{huang2021study,
  author = {Xin Huang and Hui Chen and Jing-Dong Yan},
  title = {{Study on structured method of Chinese MRI report of nasopharyngeal carcinoma}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2021},
  doi = {10.1186/s12911-021-01547-1},
}
Figure