Automated Classification of Free-Text Radiology Reports: Using Different Feature Extraction Methods to Identify...

Area of Science:

• Medical imaging analysis
• Natural Language Processing (NLP) in healthcare
• Machine learning for diagnostic support

Background:

• Radiology reports often contain unstructured free text, hindering efficient data extraction and analysis.
• Natural Language Processing (NLP) offers a solution by converting free text into machine-readable formats.
• Structured data is crucial for developing scalable and reliable data analysis methods in radiology.

Purpose of the Study:

• To classify unstructured radiograph reports for distal fibula fractures.
• To identify the most effective text mining method for this classification task.
• To evaluate the performance of different text representation and machine learning techniques.

Main Methods:

• A German language dataset of ankle radiograph reports was created and manually labeled for distal fibula fractures.
• Text representation methods including bag-of-words (BOW), TF-IDF, PCA, NMF, LDA, and doc2vec were applied.
• Extracted document vectors were used to train classifiers such as Neural Networks (NN), Support Vector Machines (SVM), and Logistic Regression (LR).

Main Results:

• The bag-of-words (BOW) method achieved the highest performance with an AUC of 0.98 and accuracy of 0.97.
• TF-IDF also showed strong results (AUC=0.97, acc=0.96), followed by NMF, PCA, LDA, and doc2vec.
• Among classifiers, Neural Networks (NN) outperformed SVM and LR, achieving an AUC of 0.91.

Conclusions:

• Automated classification of unstructured radiology reports can reliably detect distal fibula fractures.
• The bag-of-words (BOW) model is a particularly suitable feature extraction method for this application.
• NLP techniques enhance the ability to extract valuable information from free-text radiology reports.