Automated Chest X-ray Report Generation Using Attention-Enhanced GoogleNet-LSTM Architecture

Abstract

Chest radiography remains a cornerstone of clinical diagnostics, yet its interpretation is time-consuming and dependent on specialized expertise. The growing shortage of radiologists, combined with the increasing volume of imaging exams, often leads to delays and inconsistencies, highlighting the need for automated solutions. In this study, we present an automated framework for generating diagnostic reports directly from chest X-rays. The model uses GoogleNet for visual feature extraction and a long short-term memory (LSTM) network to generate reports. An attention mechanism is incorporated to focus on clinically relevant image regions. The framework was evaluated on the publicly available Indiana University (IU) Chest X-ray dataset, with performance assessed using language-based metrics (BLEU, ROUGE-L, METEOR, CIDEr) and clinical accuracy indicators, such as precision, recall, and F1-score. Results demonstrated that the attention-based architecture outperformed baseline encoder-decoder models, particularly in CIDEr and clinical F1 metrics, suggesting the reports were more fluent and clinically accurate. Attention maps showed alignment with key image areas, such as the cardiac silhouette for cardiomegaly and costophrenic angles for pleural effusion. While limitations were observed in handling rare conditions and occasional generic phrasing, the framework effectively improved the efficiency and consistency of radiology reporting.