A Comparative Study of Deep Learning Applications in Medical Report Generation Using Chest X-Rays

Abstract

The automation of chest X-ray report generation has emerged as a transformative area in medical imaging, with the potential to alleviate radiologists’ workload and enhance diagnostic accuracy and efficiency in healthcare. This paper presents a comprehensive survey of deep learning approaches aimed at automating the creation of clinical reports from chest X-rays, focusing on state-of-the-art methodologies, evaluation techniques, and real-world applicability.To offer a holistic view the survey covers key model architectures, including Convolutional Neural Networks (CNNs), Transformers, and multimodal frameworks, providing an in-depth exploration of their capabilities and limitations in medical report generation. Notably, models like ViT-GPT2 and ResNet101+Tranformer achieve promising results, with ViT-GPT2 reporting a BLEU-4 score of 0.2020 and ResNet101+Tranformer reporting a BLEU -3 Score of 0.1765 post the proposed novel report standardization. The paper examines publicly available datasets such as IU-X Ray and MIMIC-CXR, which serve as benchmarks for training and evaluating these systems, alongside commonly used assessment metrics. The paper underscores challenges inherent to these datasets, such as biases, limited clinical diversity, while also discussing strategies to address these issues.This survey highlights not only the progress made in this field but also the gaps and unresolved questions like clinical relevance and practical concerns that hinder widespread adoption. By synthesizing insights from existing studies, this work aims to guide researchers, clinicians, and developers in advancing the field of AI-driven chest X-ray report generation. It also identifies promising directions for future research and development, contributing to the broader goal of enhancing radiological practice and improving patient outcomes through trustworthy and effective AI solutions.

Ähnliche Arbeiten