AMRG

Screening mammography generates enormous reporting volume — in South Korea alone, biennial screening for women over 40 produces millions of studies a year — and the narrative report is where the interpretive work lands. Unlike a classification label, a mammography report must describe multiple views of both breasts in free text, assign a breast density category, and conclude with a BI-RADS assessment, all in language whose terminology varies between radiologists. Chest radiography has large paired image–report corpora and a mature body of report-generation work; mammography has had neither.

AMRG (Automatic Mammography Report Generation) is an end-to-end framework for the task, built by adapting MedGemma-4B-it, Google's instruction-tuned medical vision-language model, whose image tower is the medically tuned MedSigLIP encoder. Rather than retraining the backbone, the authors insert Low-Rank Adaptation (LoRA) adapters into its linear layers, so the clinical and visual-linguistic competence acquired during MedGemma's pretraining across radiology, dermatology, pathology, and ophthalmology is preserved while the model learns mammographic vocabulary and structure.

The work comes from the Research Institute and Department of Radiology at the National Cancer Center Korea and was posted as a preprint in August 2025. Its stated contribution is as much a benchmark as a model: it establishes a reproducible protocol on a public dataset for a task that previously had no standard evaluation.

#Key Features

• LoRA adaptation across the whole backbone: Adapters are injected into every linear layer — attention query, key, value, and output projections, feed-forward layers, and gated MLP up/down/output projections — spanning both encoder and decoder.
• Multi-view mammographic input: Craniocaudal and mediolateral oblique views are encoded together, so the model reasons over the standard four-image screening exam rather than a single projection.
• Purpose-built preprocessing: Otsu thresholding isolates breast tissue from background, a tight bounding box crops the region of interest, images are resized to 512×512, and left-breast images are horizontally flipped to enforce consistent laterality.
• Clinical metrics alongside language metrics: Evaluation reports BI-RADS category and breast density accuracy in addition to BLEU, ROUGE, METEOR, CIDEr, and word-level F1, since fluent text that assigns the wrong assessment category is clinically useless.
• Backbone ablation under one protocol: MedGemma-4B is compared against Qwen2.5-VL-7B, Phi-3.5-Vision, CLIP+GPT-2, and MedCLIP+GPT-2 with identical LoRA settings.

#Technical Details

Training and evaluation use DMID, a public dataset of 510 annotated mammography cases pairing high-resolution images with radiologist-written diagnostic reports, split 407 training / 51 validation / 52 test. BI-RADS categories are heavily imbalanced toward BI-RADS 1 and 3, as screening prevalence dictates. Every configuration trains for 20 epochs at batch size 4 with gradient accumulation 8, AdamW at learning rate 1e-4, and generates at temperature 0.1.

A sweep over LoRA rank r ∈ {16, 32, 64} and scaling factor α ∈ {8, 16} finds r = 32, α = 16 best overall: ROUGE-1 0.5750, ROUGE-L 0.5691, METEOR 0.6152, CIDEr 0.5818, word-level F1 0.5610, and BI-RADS accuracy 0.5582. Rank 64 consistently degrades, which the authors attribute to overfitting on a dataset this small; r = 32, α = 8 actually yields the highest BI-RADS accuracy at 0.5686. Untuned MedGemma-4B is effectively unusable on the task, scoring ROUGE-L 0.0613 and BI-RADS accuracy 0.0000. Among backbones, Qwen2.5-VL-7B comes close on surface n-gram overlap (ROUGE-L 0.5634) but falls well short on BI-RADS accuracy (0.4510) despite being larger, while Phi-3.5-Vision collapses at 0.3559 and 0.1176.

#Applications

The immediate use is draft-report generation inside a high-volume screening programme, where a system that produces a structured first pass — findings, density, BI-RADS — could shorten radiologist reporting time while a human retains sign-off. More broadly, the LoRA recipe and preprocessing pipeline are a template for adapting a general medical vision-language model to any imaging modality with a small paired image–report corpus, and the DMID protocol gives subsequent mammography work a fixed point of comparison.

#Impact

AMRG establishes the first reproducible benchmark for narrative mammography report generation, and its backbone ablation supports a useful general finding: domain-specialised pretraining beats raw parameter count on clinically specific tasks. The limits are equally clear. DMID's 510 cases and 52-case test split are small, the evaluation is confined to that single public dataset with no external multi-site validation, and BI-RADS accuracy near 56% is far from clinical deployment. The paper is a preprint, and no code repository or adapter weights have been released, so reproduction depends on reimplementing the described protocol.