FADA: Fetal Ultrasound Interpretation Dataset — 56,805 Expert Sonographer Clinical Descriptions

FADA Interpretation Dataset: Expert Sonographer Clinical Descriptions for Fetal Ultrasound This is the primary data contribution of the FADA project. An expert sonographer with >10 years of clinical experience in obstetric ultrasound imaging created structured clinical descriptions for 18,935 fetal ultrasound images (56,805 total conversations) across 14 anatomical categories. What This Dataset Contains For each fetal ultrasound image, the expert sonographer answered 8 standardized clinical questions, encoded as a structured JSON object: anatomical_structures — All visible fetal anatomical structures (e.g., "aorta, inferior vena cava, liver, stomach") fetal_orientation — Spatial orientation and presentation (e.g., "Axial upper abdomen, vertebral column to right") imaging_plane — Standard ultrasound imaging plane (e.g., "Transverse trans-abdominal plane") biometric_measurements — Obtainable biometric measurements (e.g., "AC", "BPD", "FL") gestational_age — Estimated gestational age from visible anatomy image_quality — Assessment of ultrasound image quality normality_assessment — Normal vs. abnormal findings clinical_recommendations — Suggested clinical follow-up actions Dataset Composition Total conversations: 56,805 Unique images: 18,935 Full 8-field entries: 37,870 (complete clinical assessment) 4-field subset entries: 18,935 (anatomical_structures, fetal_orientation, imaging_plane, biometric_measurements) Anatomical categories (14): Abdomen, Aorta, Cervical, Cervix, CRL-View, Femur, Non-standard NT, NT-View, Public Symphysis/Fetal Head, Standard NT, Thorax, Trans-cerebellum, Trans-thalamic, Trans-ventricular Source Images All ultrasound images are sourced from publicly available datasets released under Creative Commons Attribution 4.0 (CC-BY-4.0) licenses: FPUS23 (Fetal Pose, Zenodo) FUSEP (Fetal Brain Structures, Zenodo) HC18 / Fetal Head (Brain Segmentation, Zenodo) CRL_NT (Crown-Rump Length / Nuchal Translucency, Mendeley Data) FOCUS (First-trimester Cardiac, Zenodo) Fetal Abdominal Structures (Doppler Vessels, Zenodo) PS-FH (Pubic Symphysis-Fetal Head, Zenodo) Our contribution is the clinical interpretation annotations — the expert sonographer descriptions that transform raw ultrasound images into structured clinical knowledge. The source images themselves are not redistributed; users should obtain them from the original repositories above. File Format The dataset is provided as a JSONL file where each line contains a conversation with a user message (image path + prompt) and an assistant message (structured JSON clinical interpretation). This format is directly compatible with VLM fine-tuning frameworks (e.g., Qwen-VL, LLaVA). Validation The interpretation quality was validated by an expert sonographer on 237 images (blinded scoring, 1-3 scale) and 49 clinical cases in human-in-the-loop mode. Results: 73.5% of interpretations received a perfect score (1), with a mean score of 1.286. Full validation scoring data is included as external_validation_scoring_v2.csv. Associated Resources Paper: "FADA: Accessible fetal ultrasound interpretation and annotation with a selectively distilled unified vision-language model. Code: github.com/mahmoodphd/FADA Model Weights: huggingface.co/mshz88/FADA-SKD-4B Mobile Model (GGUF): huggingface.co/mshz88/FADA-Mobile-GGUF (Q4_K_M quantization for llama.cpp, 712 MB total) Mobile App (APK): github.com/mahmoodphd/FADA/releases (Android app for offline inference) Web Application: https://mshz88-fada-ultrasound-vlm.hf.space. Mobile Edge Deployment The FADA-SKD 0.8B model is deployed on Android via llama.cpp with GGUF quantization (Q4_K_M). Validated on Honor 90 (Snapdragon 7 Gen 1, 12 GB RAM, Android 15): full 5-phase pipeline completes in ~59 seconds entirely offline, demonstrating stand-alone integration with portable fetal ultrasound devices without cloud connectivity. Video Demonstration A video demonstration of the FADA system is available on YouTube: https://youtu.be/CbXcz74fn6k The demo showcases interactive chat mode, autonomous 5-phase analysis, and the built-in anatomy reference atlas. @misc{maani2025fetalclipvisuallanguagefoundationmodel, title={FetalCLIP: A Visual-Language Foundation Model for Fetal Ultrasound Image Analysis}, author={Fadillah Maani and Numan Saeed and Tausifa Saleem and Zaid Farooq and Hussain Alasmawi and Werner Diehl and Ameera Mohammad and Gareth Waring and Saudabi Valappi and Leanne Bricker and Mohammad Yaqub}, year={2025}, eprint={2502.14807}, archivePrefix={arXiv}, primaryClass={eess.IV}, url={https://arxiv.org/abs/2502.14807}, }