Development of Ontology-Supported and Deep Learning-Based Language Model for Chest Radiology Dataset (MIMIC-CXR)

Project no. 24110 · Status: Kapanmış (Yayın bekleniyor) · 2023

Summary

In the healthcare domain, machine learning and image‐processing techniques have been rapidly expanding in diagnostic workflows based on laboratory findings, clinical observations, magnetic resonance imaging, tomography, and radiological images. Nevertheless, many types of medical documentation—such as expert-authored epicrisis reports—contain critical and diagnostically valuable information that is not captured by imaging data alone. In this context, the field of Natural Language Processing (NLP) constitutes an essential area of research for information extraction, including the automatic generation of medical reports. By leveraging expert knowledge and clinical experience, NLP techniques enable the extraction of meaningful information from health records, thereby reducing the workload of healthcare professionals, minimizing the risk of diagnostic error, and saving time. Within this framework, the MIMIC-CXR dataset stands out as a large-scale collection of chest radiograph images paired with expert-written reports (Johnson et al., 2019). In this project, the radiology reports in the dataset underwent a comprehensive structural and semantic text analysis; extensive preprocessing techniques were applied prior to the development of any NLP-based predictive model. Syntactic and semantic analyses performed on these unstructured reports revealed the linguistic characteristics and semantic relationships embedded in the text. Subsequently, a variety of methods and language models were employed to assess the performance of clustering and classification algorithms. The performance of the proposed methods and language models on the MIMIC-CXR dataset was examined in detail, and the resulting findings were documented. Furthermore, to demonstrate the practical utility of the developed classification model in clinical settings, a web-based application was designed. This application provides both a visual and functional environment for evaluating the effectiveness of the classification model derived from the applied algorithms.The anticipated contribution of the project is to shed light on the development of NLP models tailored to the MIMIC-CXR dataset and to guide future research on AI-assisted information extraction from medical reports. In doing so, the study offers methodological insights into the design and deployment of language models and underscores the importance of integrating NLP-based approaches into health informatics for the automated analysis of medical texts.

View details