Development of mobile-deployable vision language models for layman chest radiograph summarization using parameter-efficient fine-tuning and quantization.
Authors
Affiliations (3)
Affiliations (3)
- Department of Chemical Engineering, IIT Delhi, New Delhi, Delhi, India.
- Department of Radiodiagnosis, Post Graduate Institute of Medical Education and Research, Chandigarh, India.
- Department of Chemical Engineering, IIT Delhi, New Delhi, Delhi, India. Electronic address: [email protected].
Abstract
Chest radiography is among the most widely performed diagnostic imaging procedures in clinical medicine, yet the technical reports generated by radiologists often remain incomprehensible to patients. Automated generation of patient-friendly layman summaries from chest X-ray images and radiology reports using vision-language models represents a clinically viable solution. However, the feasibility of deploying such models on consumer mobile devices remains undemonstrated, limiting their applicability in resource-limited settings and LMICs where cloud infrastructure cannot be assumed. This study aimed to investigate whether a compact vision-language model, fine-tuned using parameter-efficient methods, can be deployed on a smartphone while approaching the performance of substantially larger models. Four vision-language models of increasing size, namely Gemma3 (4B), Qwen3-VL (8B), Llama 3.2 Vision (11B), and Pixtral-2409 (12B), were evaluated under pre-trained and fine-tuned Quantized Low-Rank Adaptation (QLoRA) conditions on over 130,000 chest X-ray image-radiology-layman report triplets from the PadChest corpus. Evaluation spanned text relevance (BLEU, ROUGE, METEOR, BERTScore, semantic similarity), linguistic readability (Flesch-Kincaid Grade Level, Coleman-Liau Index, Dale-Chall Readability Score), and clinical fidelity (RadGraph-F1). Zero-shot generalization was assessed on an external MIMIC-CXR validation set, alongside a report-only ablation study. A clinical radiologist independently assessed outputs from the fine-tuned compact model using a five-point Likert scale across accuracy, completeness, conciseness, and safety criteria. Fine-tuning produced transformative gains across all models. BLEU improved by up to 12-fold, ROUGE-2 6.7-fold, and RadGraph-F1 4.8-fold. Readability shifted from a college-level baseline average of FKGL 12.72 and CLI 13.63 to consistently patient-appropriate levels averaging FKGL 7.49 and CLI 8.67 across the models. Post fine-tuning, inter-model differences remained around 0.12 on BLEU, under 0.09 on FKGL, and approximately 0.09 on RadGraph-F1. External validation preserved clinical fidelity (RadGraph-F1 0.373), and a report-only ablation confirmed the radiograph's contribution. Radiologist evaluation yielded a Likert score of 4.76/5.00, indicating promising clinical quality. The model was deployed on a consumer smartphone via GGUF quantization at a throughput of 18.26 tokens per second. Parameter-efficient fine-tuning of a compact 4-billion-parameter vision-language model yields performance approaching that of models 3 times its size for layman chest radiograph summarization. Deployment on a consumer smartphone at clinically viable inference speeds establishes on-device radiology report simplification as a practical reality.