Abstract
Indonesia carries one of the world’s largest tuberculosis (TB) burdens, while chest X-ray (CXR) screening is constrained by uneven reporting capacity and deployment resources. Vision-language models (VLMs) can generate draft radiology reports, but strong multimodal models remain costly and may miss clinically meaningful findings. We evaluate efficient Indonesian-language CXR report generation as a joint clinical-semantic and deployment problem. With evaluation on a held-out testing fold of 1,082 studies, we benchmark five candidate VLMs to identify Qwen3.5-9B as a high-cost multimodal teacher reference and CheXAgent-2-3B as an efficient student. Qwen is evaluated under FP16, INT8, and INT4 precision with native and structured-prompted inference. We then distill Qwen-to-CheXAgent using report-level knowledge distillation (KD) and joint report-representation KD. Automated TB-14 agreement is estimated using X-Raydar-derived precision, recall, and micro-F1 over the TB-oriented 14-label thoracic abnormality subset, while deployment cost is measured by latency and allocated VRAM. Prompted Qwen INT8 achieved the highest Qwen micro-F1 (0.484) but was slow, whereas prompted Qwen INT4 retained 0.446 micro-F1 while reducing allocated VRAM by 53.4% and latency by 23.1% versus FP16. The proposed RRKD student improved native CheXAgent from 0.376 to 0.411 micro-F1 with nearly unchanged allocated VRAM and improved several TB-relevant findings. These results support efficiency-aware CXR report generation as an efficiency–quality frontier rather than compression alone.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMA_026.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{CahRet_Accelerating_MICCAISAT2026,
author = { Cahyaningsih, Retno Dwi AND Yunus, Reyhan Eddy AND Astagiri, Prasandhya Yusuf AND Montana, Giovanni AND Valindria, Vanya},
title = { { Accelerating Chest X-ray Report Generation for Tuberculosis Screening in Indonesia: An Efficiency-Aware Benchmark of Quantized and Distilled Vision-Language Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}