Kaggle Med-Gemma Impact Challenge:

Theodore’s Heart Initiative is pleased to share an update regarding our participation in the Kaggle Med-Gemma Impact Challenge. Kaggle is an internationally recognized machine learning competition platform that has been hosting cutting edge challenges to solve the worlds most complex computational problems for years. Our team is currently developing an application to enter in the competition which is designed designed to support medical professionals in the early detection of congenital heart defects (CHDs) in support of our broader mission.

The Objective: Supporting the Routine Anatomy Scan

The mid-trimester ultrasound is a standard part of prenatal care, yet detecting heart defects remains a significant clinical challenge. The fetal heart is a small, rapidly moving target, and obtaining the specific anatomical views required for a diagnosis requires high precision from the sonographer.

The Technical Framework

The application is designed as a two-stage digital assistant that integrates into the existing clinical workflow.

Stage 1: Real-Time Quality Gate & Navigation: This module monitors the live ultrasound feed to identify when the heart is in the correct position. To ensure a comprehensive assessment, the application guides the technician to capture the five standard anatomical planes required by international guidelines.

• Abdominal Situs: Confirms the stomach and heart are on the correct side.

• Four-Chamber View (4CH): Evaluates the overall symmetry and basic structure of the heart.

• Left Ventricular Outflow Tract (LVOT): Checks the aorta’s connection to the heart.

• Right Ventricular Outflow Tract (RVOT): Checks the pulmonary artery's connection.

• Three-Vessel Trachea View (3VT): Assesses the alignment and size of the major blood vessels.

  
  

Once the application confirms an image meets a high-fidelity "Gold Standard" threshold, it provides a visual green-border indicator to the technician.

Stage 2: Clinical Reasoning Engine. High-quality images are then analyzed by a specialized medical reasoning engine. This system evaluates structural integrity, such as the heart walls and chamber balance, providing a written summary of its observations to assist the final medical review.

Impact on Clinical Outcomes

By providing real-time feedback, the application aims to assist human operators in a way that can lead to more consistent results. By focusing on standardization, the tool helps ensure that every scan captures the same set of high-quality images, regardless of clinical setting. Improved efficiency through automated view recognition can reduce the time and mental effort required to navigate complex scans. Integration early into the existing human workflow provides the maximum exposure of the tool to as many families as possible. This is critical as it allows the process to leverage the existing expertise of the medical staff by surfacing issues that need a closer look with specialty tools that may not previously have been able to be identified early enough to make the difference.

Data Utilization

Our current phase of development leverages the Fetal Heart Four-Chamber Ultrasound Dataset, which provides expert-labeled images for training and validation. This ensures that our assistant's "Quality Gate" is calibrated against established medical benchmarks. Our goal is to empower sonographers with tools that reduce the technical difficulty of CHD screening, ultimately helping more families receive timely and accurate diagnoses.

Want to explore more? Dive into the experiment details and discover fascinating insights.