Abstract

Ultrasound report generation (US-RG) can ease sonographers’ documentation burden, but it is a challenging task. Practical clinical deployment requires US-RG lightweight models, but progress is constrained by the scarcity of open-source ultrasound image-text datasets. Under these low-data and resource constrained conditions, we study how the choice of a lightweight US-RG model’s pretrained initialisation and component-level (vision/language model) finetuning affects how it performs on multi-anatomy datasets. Our findings suggest that: (i) domain specific ultrasound pretraining may not be necessary for lightweight US-RG as general biomedical initialisation and an initialisation from web-scale pretraining perform comparably; (ii) given the limited data, scaling the model yields little gains, whereas parameter-efficient adaptation of both the vision and language model is more promising; (iii) task-specific finetuning irrespective of the initialisation source weights outperforms zero-shot evaluation and (iv) lightweight models suitable for on-device deployment can be used for US-RG, but the gap between the semantic similarity and accurate identification of clinically relevant findings shows the need for methods that improve clinical factuality.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_089.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=Xor2PSftMt

BibTex

@InProceedings{MloSha_Effect_MICCAISAT2026,
        author = { Mlotshwa, Sharon AND Saha, Pramit AND Bentahar, Jamal AND Naseer, Muzammal AND Alsharid, Mohammad},
        title = { { Effect of Pretrained Initialisation and Component-Wise Tuning on Lightweight VLMs for Ultrasound Report Generation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17264},
        month = {pending},
        page = {pending}
}


back to top