CHEST XRAY AUTOMATED REPORT GENERATION

Abstract

This thesis uses NLP and deep learning to automate chest X-ray image textual reports to improve diagnostic precision and efficiency. The best encoder-decoder architecture is Vision Transformer (ViT) and Bidirectional and Auto-Regressive Transformers (BART), according to the research. The distinguishing feature of ViT is its self-attention mechanism, enabling it to effectively capture distant relationships within an image. The system analyzes images by dividing them into non-overlapping patches and considers each patch as a token in natural language processing (NLP) models. This method enables thorough examination of visual characteristics and detection of subtle abnormalities, which are crucial in the field of medical imaging. BART is a variation of transformer and is trained on a large amount of text data. It combines bidirectional encoders and autoregressive decoders, which allows it to generate text that is both coherent and contextually relevant. The ViT-BART model combines the features of the Vision Transformer (ViT) and BART models to extract detailed embeddings from chest X-ray images. These embeddings are then passed through BART’s encoder to create a latent representation that contains contextual information. The experimental findings demonstrate that the ViT-BART model surpasses other examined architectures, thereby improving the precision of automated medical report generation, interpretability, and clinical relevance. This research has the capacity to greatly influence clinical workflows and improve patient outcomes.