Effective Natural Language Processing and Interpretable Machine Learning for Structuring CT Liver-Tumor Reports

Abstract

In the past, the liver tumors were reported manually in an unstructured format. There actually exists much valuable knowledge in these reports for further disease risk assessment, disease recognition and treatment recommendation. Yet, it is not easy to read and mine knowledge from the unstructured reports. Hence, how to extract the knowledge from these biomedical reports effectively and efficiently has been a challenging issue in the past decades. Although a set of Natural Language Processing techniques were proposed for Bio-medical information retrieval, few related works were made on transforming the unstructured CT liver-tumor reports into structured ones. To aim at this issue, in this paper, we propose a two-stage report structuring method by integrating effective Natural Language Processing (NLP) and interpretable machine learning. For the first stage, the candidate keywords in unstructured reports are extracted. Next, the feature keywords are determined by the feature-selection technique. For the second stage, the well-known multi-classifiers are performed, and finally the reports are labeled in a refined structure format. Further, the factor keywords in the classification model are filtered to interpret the performance. In overall, the proposed report structuring method generates a hierarchical data structure, including the common features and refined features in the 1st and 2nd levels/stages, respectively. To reveal the performance of proposed method, a set of evaluations were conducted and the results show that, the proposed method is more promising than the fashion neural networks such as Bert (Bidirectional Encoder Representations from Transformers) in terms of effectiveness and efficiency.