Research map / Vision-language and multimodal / Multimodal LLMs

F$^2$-Assist: Multi-Phase Fetal Growth Forecast and Report Generation from Ultrasound Examination

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces multi-phase fetal growth forecasting and report generation from sequential ultrasound, releases the GrowthFetus dataset (9,280 exams, 2,000 fetuses), and proposes F2-Assist, an MLLM with organ alignment, temporal encoding, and a numeric soft-token adapter.

Problem

Existing medical MLLMs for ultrasound report generation handle single examinations and qualitative description only, ignoring longitudinal history and precise continuous biometric values needed to forecast fetal growth.

Approach

A unified MLLM combining query-driven cross-phase organ prototyping, causal history-aware temporal transformer over phases with gestational-age positional embeddings, and a Growth Parameter Adapter that embeds z-normalized biometry as soft tokens plus a regression head trained with Smooth-L1.

Key results

• F2-Assist (8B) reaches R2 0.78 and MAE 0.80 cm on next-phase biometric prediction, versus 0.59 R2 for the best medical MLLM baseline (Lingshu, MAE 1.13).
• On report generation, F2-Assist achieves B4 0.53 and CIDEr 3.66, the best among compared MLLMs.
• Ablation: removing Cross-Phase Organ Alignment drops R2 from 0.78 to 0.35 and raises MAE from 0.80 to 0.93.

Surprising findings

• Digit-as-text encoding of numbers in the base Qwen2.5-VL performs worse than the numeric adapter, and removing numeric history inputs drops R2 to 0.29.

Limitations

• Accuracy for AC and FL degrades beyond 36 weeks, attributed to sparse late-gestation training data.
• Performance drops with history longer than four phases, attributed to data sparsity and noise.
• Compute is described only as A100 GPUs without a count; the dataset is not stated to be public.

Experimental setup

Datasets: GrowthFetus

Benchmarks: GrowthFetus

Metrics: BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L CIDEr MAE Accuracy@±5% Accuracy@±10%

Baselines: LLaVA-1.5 LLaVA-Med R2GenGPT Qwen2.5-VL InternVL-3 HuatuoGPT Lingshu Growth-Chart Transformer time-series LSTM

Builds on: LLaVA Qwen2.5-VL CLIP LoRA

Closest papers by meaning

• FETAL-GAUGE: A BENCHMARK FOR ASSESSING VISION-LANGUAGE MODELS IN FETAL ULTRASOUND (ICLR 2026 · similarity 0.89)
• U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding (ICLR 2026 · similarity 0.85)
• MCAT: Visual Query-Based Localization of Standard Anatomical Clips in Fetal Ultrasound Videos Using Multi-Tier Class-Aware Token Transformer (AAAI 2025 · similarity 0.85)
• Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images (ICML 2024 · similarity 0.83)
• BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models (CVPR 2026 · similarity 0.83)
• Toward a Vision-Language Foundation Model for Medical Data: Multimodal Dataset and Benchmarks for Vietnamese PET/CT Report Generation (NeurIPS 2025 · similarity 0.82)