Knowledge graph-based small sample learning for label of medical imaging reports

Abstract

Background: Through the utilization of knowledge graph and small sample learning, the study effectively tackled the challenges of data scarcity and automatic annotation in the field of medical image recognition with the application of artificial intelligence technology. Methods: Initially, 2000 X-ray reports of the lumbar spine were labeled manually employing a knowledge graph approach. These reports were subsequently split into a training dataset of 1000 cases and a test dataset of 1000 cases. Following this, distinct levels of data augmentation, namely the synonym/apposition method, were applied to the training dataset. Subsequently, the deep learning model BERT (Bidirectional Encoder Representation of Transformer) was utilized for the training process. Afterward, the BERT model is tested on the specified test dataset, and subsequently, the nodes showing insufficient performance are supplemented with iterative target data. Finally, the method is evaluated by using various metrics including AUC(Area Under Curve), F1 score, precision, recall and relabelled rate. Results: Before conducting data augmentation, the AUC value was 0.621, the F1 value was 32.1%, the average precision was 0.383, and the average recall was 0.303. Following data augmentation, the AUC value improved to 0.789, the F1 value improved to 70.3%, the average precision improved to 0.879, and the average recall improved to 0.580. After targeted data supplementation, the AUC reached 0.899, the F1 value reached 85.7%, the average precision reached 0.952, and the average recall reached 0.803. Conclusions: The current study achieves its objective by training an automatic annotation model using a knowledge graph-based approach to annotate medical imaging reports on a small sample dataset. Furthermore, this approach enhances both the efficiency and accuracy of medical imaging data annotation, providing a significant research strategy for applying artificial intelligence in the field of medical image recognition.