Abstract
Structured report generation improves diagnosis and reduces the inter-radiologist interpretation variability; however it is quite challenging and needs clinical understanding. The design of structured clinical schemas through parameter-efficient adaptation of Vision Language Models (VLMs) remains unexplored. This paper revolves around the systematic study of parameter efficient fine tuning using Low Rank Adaptation (LoRA) applied to four state-of-the-art Vision language Models (VLMs) including Med-Gemma-4b, Qwen2.5-VL-7B, Bio-Mistral and BLIP-2, to generate structured mammography reports using Digital Mammography Image Dataset (DMID). Our framework enforces a standardized instruction prompt and template with multimodal inputs processed in a unified token sequence to generate structured report covering breast composition, findings, impression, BI-RADS category and breast density. Reports are linguistically evaluated using the standard NLP metrics (BLEU-n score, ROUGE, METEOR, and CIDEr) alongside clinical attribute accuracy and prompt sensitivity analysis to evaluate behavior of VLMs towards different prompts. With only 0.14% trainable parameters, our approach achieves BLEU-1: 0.6503, ROUGE-L: 0.7111, CIDEr: 1.5703 and Density accuracy: 0.5000. Wilcoxon test statistically ensures that Med-Gemma-4B with LoRA integration at the attention projection layers significantly outperforms all competing models. The results demonstrates that parameter efficient fine tuning produce structured report generation with clinical relevance, substantially advancing over prior unstructured generation approaches.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_029.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=d0DCEKBvrn
BibTex
@InProceedings{ShaAle_Structured_MICCAISAT2026,
author = { Shafiq, Aleena AND Jadoon, Khurram AND Shah, Yasir Ali},
title = { { Structured Report Generation from Digital Mammograms via Instruction-Tuned LoRA Adaptation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}