Fine-tuned BERT Language Model for Efficient Nuclear Medicine Data Retrieval

Abstract

Efficient data retrieval is crucial in medical databases, where quick access to relevant patient information can significantly impact data management. In this work, we propose an Natural language processing (NLP)-based approach for predicting the type of scan from nuclear medicine reports to enable efficient data retrieval. We utilized a deep learning approach, using the fine-tuned BERT model, to classify the scan type based on the finding. The scan types we predicted include 51 different classes, such as MPI SPECT, Whole-Body Bone Scans, Thyroid Scans, and Lung Perfusion. To train and evaluate the model, we used a 75/10/15 split for the training, evaluation, and test phases, respectively. The radiology reports were generated by different nuclear medicine physicians with varying years of experience, ensuring a high variability of the dataset. We evaluated our model on various types of scans and achieved an accuracy of 98% on a dataset consisting of 36,000 nuclear medicine reports from different centers. Our proposed approach has the potential to significantly improve the efficiency of data retrieval from medical databases by automating the process of extracting scan names from radiology reports. This will enable faster and more accurate retrieval of relevant information for medical professionals, thus improving the quality of patient care.

Ähnliche Arbeiten