Anandi: A Multimodal AI-Based System for Automated Fetal Biometry and Intelligent Decision Support

Anandi is an intelligent auxiliary system for prenatal medical examinations, integrating technologies such as the DETR computer vision model, fine-tuned Gemma large language model, LangGraph state management, and RAG (Retrieval-Augmented Generation) with LanceDB vector database. Its core functions are automatic biometric measurement of fetal ultrasound images and intelligent filling of PC-PNDT Form F, aiming to improve efficiency and accuracy, reduce the burden on medical staff, and ensure compliance.

Traditional fetal biometry relies on manual operation of ultrasound equipment for readings, which is time-consuming and prone to errors; Indian law requires prenatal diagnosis institutions to submit PC-PNDT Form F, and manual filling is cumbersome, easy to miss errors, and has compliance risks.

• Computer Vision: Based on the DETR end-to-end object detection model, no anchor boxes or non-maximum suppression required, directly analyzing ultrasound images to extract measurement data;
• NLP: Using a fine-tuned Gemma 4B model to understand medical terminology and clinical context;
• State Management: Coordinating component interactions via LangGraph to manage the complete process from image input to report generation;
• Knowledge Enhancement: RAG architecture combined with LanceDB vector database to retrieve external knowledge such as medical literature/guidelines to overcome model hallucinations;
• System Design: Frontend-backend separation architecture, where the backend handles AI inference and business logic, and the frontend processes user interactions.

• Clinical Value: Automatic biometry improves examination efficiency and accuracy; intelligent filling of PC-PNDT forms ensures document completeness and standardization;
• Resource Support: An online demo environment is provided (anandi-ai.vercel.app), developed using Python for easy understanding and contribution by developers.

Anandi is a typical paradigm of medical AI application in vertical fields, demonstrating the idea of organically integrating multiple technologies to solve clinical problems; the technology selection (DETR/Gemma/LangGraph/LanceDB) has reference value; the balance between technological advancement, compliance, and practicality, as well as the user-centric design concept, are worthy of recognition.