Driven by deep learning advancements, automated medical imaging report generation—an essential tool for radiologists—has become a core research direction in medical AI. Existing methods predominantly construct end-to-end models based on forward sequential workflows. Despite phase-specific progress, these methods have an inherent flaw: unidirectional information flow, where subsequent modules passively adjust to preceding outputs, and vice versa. This cross-module information disconnect often leads to notable discrepancies between generated reports and actual clinical diagnoses. Inspired by visual-neural interaction mechanisms, this study proposes a Feedback-Attentive Multiscale Network (FAMR-Net), which regulates the entire process of radiology report generation through a designed crosshierarchical dynamic feedback mechanism. Within this framework, each hierarchical level and functional module supports bidirectional output regulation, enabling dynamic coordination of cross-module information flow. Furthermore, to enhance the precision of semantic expression in clinical text, a dual-track word embedding strategy is designed to improve the clinical relevance of generated reports. Experiments on the two benchmark datasets (IU-Xray and MIMIC-CXR) demonstrate that the proposed method outperforms state-of-the-art models across multiple key metrics, with BLEU-1 scores reaching 0.505 and 0.397, and ROUGE scores achieving 0.410 and 0.297, respectively. Ablation experiments further validate the necessity and effectiveness of each core component of the framework. This research provides crucial technical support for alleviating radiologists’ burden of report composition and advancing the automation of clinical diagnostic workflows, holding significant clinical application value.
Zuo et al. (Mon,) studied this question.