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

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Medical Diagnostics

Background:

• Cervical optical coherence tomography (OCT) provides high-resolution tissue visualization but faces diagnostic challenges with limited supervision.
• Vision-language models (VLMs) show promise but struggle with artifacts that mimic biological structures, hindering pathological analysis.
• Artifacts in OCT images can distract models, obscuring crucial details of layer degradation essential for accurate diagnosis.

Purpose of the Study:

• To introduce SAVLT, a structure-aware tuning framework for adapting VLMs in cervical OCT analysis.
• To enhance diagnostic reliability by addressing confounding artifacts and improving focus on tissue pathology.
• To enable robust few-shot generalization and clinical interpretability for foundation models in heterogeneous OCT imaging.

Main Methods:

• SAVLT employs parameter-efficient fine-tuning to adapt VLMs.
• A region-aware spatial attention (RaSA) module is introduced to enforce spatial constraints and purify visual representations.
• A dual-constraint objective combines image-text alignment with visual prototypes for stable optimization across domains.

Main Results:

• SAVLT effectively shifts VLM attention from global matching to anatomical grounding.
• The RaSA module successfully removes non-biological noise, restoring focus on intra-tissue structural integrity.
• The framework demonstrated robust few-shot generalization and clinical interpretability across multi-center datasets.

Conclusions:

• SAVLT establishes a reliable paradigm for deploying foundation models in cervical OCT imaging.
• The proposed structure-aware tuning framework significantly improves diagnostic accuracy by mitigating artifact interference.
• SAVLT offers a promising solution for trustworthy AI-driven diagnostics in medical imaging with limited supervision.