The Power of Context: A Comparative Analysis of Text Embeddings in Radiology Report Generation

In radiology report generation, the representation of medical language is crucial for producing accurate and meaningful descriptions of imaging findings. However, accurately capturing and representing complex medical language and imaging findings remains a significant challenge. The variability in radiology report structures, specialized medical terminology, and the need for precise contextual understanding complicate the automated generation of high-quality reports. While recent advancements in natural language processing (NLP) have introduced powerful text embeddings, it remains unclear which of these methods best addresses the unique demands of radiological language. This study aims to address this gap by conducting a comparative analysis of several prominent text embedding techniques, including BERT, BioBERT, clinicalBERT, and BioGPT, to understand their effectiveness in generating radiology reports. We employ a framework that integrates image features extracted from Vision Transformer (ViT) with contextualized text embeddings, followed by Long Short Term Memory Network (LSTM) decoder and Generative Pre-Trained Transformer 2 (GPT -2) decoder for report generation. Our experiments reveal that BioGPT outperforms other embeddings, demonstrating a superior ability to capture complex medical concepts and provide coherent description up to 20% better than other embedding models in both LSTM and GPT-2 decoder. These findings highlight the importance of selecting appropriate text embeddings for optimizing the quality and clinical relevance of automated radiology report generation.