From Images to Insights: The Role of AI in Radiology Report Generation

Radiology reports are important for diagnosing and managing a variety of health conditions, but manually crafting them is always laborious and often tiresome for radiologists. Machine learning has made significant advances and is beginning to show promise to help automate this process in recent years by enabling the production of reports from a clinical image, such as a chest X-ray or CT scan. This paper provides a review of the state of machine learning in Radiology Report Generation. We focus on the ability to use Convolutional Neural Networks (CNNs) to understand image data as well as transformer-based models to produce text data. In addition, we will examine the newer multimodal models which produce output combining image and text for potentially more accurate and contextually aware reports. Furthermore, our review highlights important observations, particularly the growing use of Deep Learning models and large medical datasets (e.g., MIMIC-CXR, CheXpert). Although there has been advancement, challenges remain such as clinically inaccurate results, lack of explainability, and inadequate variety of data. In summary, we conclude that although machine learning has a strong potential to significantly reduce work for radiologist and add consistent reporting, substantial additional work needs to be done to bring these systems to greater trust and reliability for real-world clinical settings.