Evaluating IndoBERT Deep Learning NLP for Disease Classification in Radiology Reports

Radiology has now become a primary diagnostic tool for many diseases and plays an important role in monitoring treatment and predicting outcomes. A radiologist typically creates a text report to accompany the images produced by this medical technology. These reports are often long and complex. On average, one case can take 10 minutes to interpret. The manual interpretation of a large number of images makes this process highly susceptible to human error. On the other hand, advancements in the field of artificial intelligence (AI), particularly in natural language processing (NLP), have made significant contributions to the analysis of textual data across various domains. In this study, we demonstrate IndoBERT’s performance as one of the NLP methods specifically trained on a dataset of the Indonesian language corpus to perform multiclass classification tasks in automatically predicting diseases from the textual descriptions of X-ray images written in radiology reports. The dataset used in this study were collected from several hospitals in Jakarta, Indonesia. It includes 1,719 reports from many body parts X-ray images written in Indonesian language. After the text extraction process, we have 24 different disease types from the dataset that we used as the label for our model. The IndoBERT model’s performance shows very satisfactory result with an accuracy of 93% and a macro F1-score of 88%. This study demonstrates the effectiveness of IndoBERT in enhancing diagnostic efficiency which will be beneficial for reducing the human error risk during the process of writing the report.