TGIAlign: Text-guided dual-branch bidirectional framework for cross-modal semantic alignment in medical...

Medical image-text alignment remains challenging due to subtle lesion patterns, heterogeneous vision-language semantics, and the lack of lesion-aware guidance during visual encoding. Existing methods typically introduce textual information only after visual features have been computed, leaving early and mid-level representations insufficiently conditioned on diagnostic semantics. This limits the model's ability to capture fine-grained abnormalities and maintain stable alignment across heterogeneous chest X-ray datasets. To address these limitations, we propose TGIAlign, a text-guided dual-branch bidirectional alignment framework that applies structured, lesion-centric cues to intermediate visual representations obtained from the frozen encoder. A large language model (LLM) is used to extract normalized, attribute-based lesion descriptions, providing consistent semantic guidance across samples. These cues are incorporated through the Text-Guided Image Feature Weighting (TGIF) module, which reweights intermediate feature outputs using similarity-derived weights, enabling multi-scale semantic conditioning without modifying the frozen backbone. To capture complementary visual cues, TGIAlign integrates multi-scale text-guided features with high-level visual representations through a Dual-Branch Bidirectional Alignment (DBBA) mechanism. Experiments on six public chest X-ray datasets demonstrate that TGIAlign achieves stable top-K retrieval and reliable text-guided lesion localization, highlighting the effectiveness of early semantic conditioning combined with dual-branch alignment for improving medical vision-language correspondence within chest X-ray settings.