CXRG-SVLM: a parameter-efficient small vision-language model for multi-view chest X-ray report generation in resource-constrained environments
Abstract
Automated chest X-ray report generation reduces reporting delays in resource-constrained clinical environments. Existing vision-language models often rely on large language models or full fine-tuning. This dependence causes high computational costs and limits on-premise deployment. We propose CXRG-SVLM. This small vision-language model combines a frozen RAD-DINO medical image encoder with a 4-bit quantized Qwen2.5-3B language model adapted using 4-bit QLoRA. A multi-view fusion module integrates frontal and lateral radiographs into a compact continuous visual representation. The resulting visual tokens are projected into the language-model embedding space and concatenated with clinical-indication embeddings to form a conditioning prefix. The decoder uses its native causal self-attention over this prefix and the generated tokens to produce the Findings report. The assembled implementation reports approximately 1.99 billion loaded parameter elements, of which approximately 38.1 million are trainable, corresponding to 1.91% of this framework-reported inventory. We evaluated the model on the Indiana University Chest X-ray dataset to generate the Findings section of the radiology report. The proposed model attains a ROUGE-L score of 0.3527 and a METEOR score of 0.2744. Automated entity-level agreement varies across findings, with the strongest F1 for normal descriptors and measurable agreement for pleural effusion, pneumothorax, and mediastinal mentions under the reported rule-based evaluation. Efficiency testing shows an average generation time of 5.27 seconds per report on a single graphics processing unit. These results establish the model’s computational feasibility and motivate subsequent radiologist evaluation of factual correctness, safety, and practical usefulness. Ultimately, CXRG-SVLM demonstrates that a parameter-efficient approach with less than 2% trainable parameters maintains competitive semantic fidelity, establishing an accessible foundation for future, prospectively validated on-premise clinical decision-support tools.
Related
- Fine-tuning large language models in resource-constrained environments: Methods and trade-offs
- MI-PEFT: Mixture-of-Experts Integrated Parameter-Efficient Fine-Tuning Protein Language Models Improves Acidophilic Proteins Classification
- Evaluating parameter-efficient fine-tuning of large language models for sentiment analysis with bayesian optimization
- ConLoRA-Sat: Parameter-Efficient Synthetic Satellite Imagery Generation for Data-Scarce Humanitarian Contexts
- FedEpitope: Parameter-Efficient Federated Protein Language Models for Linear B-Cell Epitope Prediction
BibTeX
@article{fareed2026cxrgsvlm,
title = {{CXRG-SVLM: a parameter-efficient small vision-language model for multi-view chest X-ray report generation in resource-constrained environments}},
author = {Muhammad Fareed and Jamal Uddin and Abid Hussain Sayyid and Adeel Ahmed and Muhammad Awais Sattar},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-77203-3},
url = {https://www.nature.com/articles/s41598-026-77203-3},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - CXRG-SVLM: a parameter-efficient small vision-language model for multi-view chest X-ray report generation in resource-constrained environments
AU - Muhammad Fareed
AU - Jamal Uddin
AU - Abid Hussain Sayyid
AU - Adeel Ahmed
AU - Muhammad Awais Sattar
PY - 2026
DA - 2026/10/29
JO - Scientific Reports
DO - 10.1038/s41598-026-77203-3
UR - https://www.nature.com/articles/s41598-026-77203-3
AB - Automated chest X-ray report generation reduces reporting delays in resource-constrained clinical environments. Existing vision-language models often rely on large language models or full fine-tuning. This dependence causes high computational costs and limits on-premise deployment. We propose CXRG-SVLM. This small vision-language model combines a frozen RAD-DINO medical image encoder with a 4-bit quantized Qwen2.5-3B language model adapted using 4-bit QLoRA. A multi-view fusion module integrates frontal and lateral radiographs into a compact continuous visual representation. The resulting visual tokens are projected into the language-model embedding space and concatenated with clinical-indication embeddings to form a conditioning prefix. The decoder uses its native causal self-attention over this prefix and the generated tokens to produce the Findings report. The assembled implementation reports approximately 1.99 billion loaded parameter elements, of which approximately 38.1 million are trainable, corresponding to 1.91% of this framework-reported inventory. We evaluated the model on the Indiana University Chest X-ray dataset to generate the Findings section of the radiology report. The proposed model attains a ROUGE-L score of 0.3527 and a METEOR score of 0.2744. Automated entity-level agreement varies across findings, with the strongest F1 for normal descriptors and measurable agreement for pleural effusion, pneumothorax, and mediastinal mentions under the reported rule-based evaluation. Efficiency testing shows an average generation time of 5.27 seconds per report on a single graphics processing unit. These results establish the model’s computational feasibility and motivate subsequent radiologist evaluation of factual correctness, safety, and practical usefulness. Ultimately, CXRG-SVLM demonstrates that a parameter-efficient approach with less than 2% trainable parameters maintains competitive semantic fidelity, establishing an accessible foundation for future, prospectively validated on-premise clinical decision-support tools.
SN - 2045-2322
ER -