SAVLT: Structure-Aware Vision-Language Tuning for Multi-Center Cervical OCT Diagnosis.

Abstract

Cervical optical coherence tomography (OCT) enables micrometer-scale visualization of tissue, yet trustworthy diagnosis under limited supervision remains challenging. While vision-language models (VLMs) offer a solution by leveraging consistent clinical semantics, their adaptation is hindered by confounding artifacts that masquerade as biological structures. This phenomenon hijacks global attention, obscuring the pathological layer degradation that is crucial for diagnosis. To overcome this limitation, we propose SAVLT, a structure-aware tuning framework that adapts VLMs via parameter-efficient fine-tuning. To shift from global matching to anatomical grounding, SAVLT introduces a region-aware spatial attention (RaSA) module that enforces spatial constraints. RaSA purifies visual representations from non-biological noise, restoring the model's focus on intra-tissue structural integrity. Furthermore, a dual-constraint objective couples image-text alignment with learnable visual prototypes to stabilize optimization against multi-center domain shifts and prompt variations. Validated across a multi-center cohort and two external test datasets, SAVLT delivers robust few-shot generalization and clinical interpretability, establishing a reliable paradigm for deploying foundation models in heterogeneous OCT imaging. Source code is publicly available at https://github.com/rabbit-my/SAVLT.