Visual Grounding and Explainability for Prompt-Driven Radiology Report Generation

Vision-language models can help automate radiology report generation, but generated reports do not always accurately reflect abnormalities visible in an X-ray. We propose a parameter-efficient framework combining a frozen RAD-DINO vision encoder, a spatial adapter, and a 4-bit quantized Qwen2.5-3B language model with LoRA adaptation for report generation and visual grounding. We evaluate report generation on IU X-Ray and report generation and grounding on an abnormal-only subset of PadChest-GR. We compare localization with Grad-CAM, HiResCAM, and Score-CAM, alongside simple spatial baselines. Results show that the framework can generate reports and syntactically valid bounding boxes with few trainable parameters; however, substantial report under-generation and weak visual alignment remain. Simple spatial priors also outperform some CAM methods on localization metrics, highlighting the importance of appropriate baselines when evaluating visual explanations. Our findings demonstrate that plausible radiological text and valid grounding syntax do not necessarily imply faithful image-specific grounding.