Performance of Vision-Language Models for Zero-Shot Lung Nodule Detection on Chest Radiographs

Background and Objectives: The ability of vision-language models (VLMs) to detect lung nodules on chest radiographs remains uncertain. This retrospective study aimed to compare the zero-shot performances of six VLMs for lung nodule detection using data from the Japanese Society of Radiological Technology (JSRT) chest radiograph database. Methods: A total of 247 chest radiographs from the JSRT database (154 with nodules and 93 without) were preprocessed and evaluated using six VLMs: RadVLM, gpt-4o-mini, Qwen3-VL-8B-Instruct, MedGemma-4b-it, LLaVA-Rad, and CheXpert Plus Model. Each model was tested using a zero-shot setting. The text outputs were binarized into nodule-present or nodule-absent labels by consensus between the two radiologists. Sensitivity, specificity, accuracy, precision, and F1 scores were calculated. Pairwise differences in sensitivity, specificity, and accuracy were assessed using McNemar test with Holm correction. Results: The overall performance was limited across all models. RadVLM achieved the highest accuracy (44.5%, 110/247) with perfect specificity (100.0%, 93/93) and precision (100.0%); however, its sensitivity was low (11.0%, 17/154). LLaVA-Rad showed the highest sensitivity (27.3%, 42/154) and F1 score (37.7%), but lower specificity (71.0%, 66/93). MedGemma-4b-it achieved 100.0% specificity, with a sensitivity of only 5.2% (8/154). Grade-specific analysis showed that detection rates were highest for obvious nodules and remained limited for subtle nodules. Pairwise analyses revealed significant differences in sensitivity and specificity for the selected model pairs, particularly between RadVLM and LLaVA-Rad. Conclusion: Current VLMs show limited zero-shot generalizability for lung nodule detection in the JSRT database, with marked trade-offs between sensitivity and specificity. Their near-term value may lie more in radiologist-assisted workflows than in stand-alone detection. Clinical Impact: Current VLMs should not be used as stand-alone tools for lung nodule detection on chest radiographs because of their limited sensitivity and substantial model-dependent trade-offs. However, their high-specificity outputs in some models and higher-sensitivity behavior in others suggest potential roles in radiologist-assisted workflows, such as report drafting and second-reader support.