Ultrasound AI Inference

3D TEE cardiac-valve view automation

@ Samsung Medison

Medical AIPythonC++Segmentation

Overview

Built during my AI & SW internship at Samsung Medison: an inference pipeline for the MVQ (Mitral Valve Quantification) workflow that automatically recognizes and aligns heart-valve structures in 3D transesophageal echocardiography.

Gallery

MVQ workflow design — MVQ workflow - from 3D TEE input to Surgeon's View

AI inference performance graph — Inference speed improvement (6s → 4s)

Samsung Medison ultrasound system — Target ultrasound platform

Highlights

Surgeon's View automation

Computes anatomical landmark points from segmentation output to auto-generate and rotation-correct the Surgeon's View - identical output regardless of input orientation.

Python → C++ production port

Prototyped model inference and visualization in Python, then ported to C++ for engine integration - improving inference speed 1.5× (6s → 4s).

Clinician-driven requirements

Requirements came directly from clinician interviews; the delivered module improved scan consistency and cut setup time by 35%.

Stack

PythonC++PyTorch3D SegmentationMedical Imaging