Intouch Chunlakittiphan

Khon Kaen University,
Thailand

Abstract Title: AI-Powered Classification of Prosthetic Heart Valves using State-of-the-Art YOLOv26: Detection and Brand Identification of Aortic, Mitral and Tricuspid Prostheses

Biography:

Intouch Chunlakittiphan, MD, is a cardiothoracic surgeon based in Khon Kaen, Thailand, currently practicing at Srinagarind Hospital and the Queen Sirikit Heart Center. Born on January 8, 1997, he has a strong clinical and academic interest in advancing cardiothoracic surgery through the integration of artificial intelligence. His work focuses on improving diagnostic accuracy and perioperative outcomes, particularly through innovative applications of AI in cardiovascular imaging and decision-making. Dr. Chunlakittiphan is committed to contributing to the evolving landscape of cardiac surgery through research, education, and international collaboration.

Research Interest:

Cardiology is rapidly advancing with artificial intelligence. This study developed three separate object detection models using YOLOv26 — one of the current state-of-the-art real-time detection models — to automatically localize and classify different brands of prosthetic heart valves from plain chest radiographs. Its lightweight and efficient architecture allows easy real-world deployment on standard hospital computers, laptops, or portable devices without needing expensive hardware.

Individual models were trained for three valve positions using real clinical images:

• Aortic valve model (8 brands: Perimount, SJM, ATS, On-X, Trifecta, Inspiris, Solo Smart, Hancock II) — 187 test images.
• Mitral valve model (9 brands: ATS, Physio, Hancock II, Perimount, On-X, CG Future, SJM, Profile 3D, Cosgrove) — 109 test images.
• Tricuspid valve model (3 brands: Contour 3D, MC3, Hancock II) — 14 test images.

Datasets were split at patient level (80/10/10) with stratified brand balancing. At the optimal threshold of 0.2, the models achieved:
• Aortic: Accuracy 96.61%, Precision 0.5358, Recall 0.5133, Specificity 0.9775, F1-score 0.5200
• Tricuspid: Accuracy 92.85%, Precision 1.0000, Recall 0.7500, Specificity 1.0000, F1-score 0.8571
• Mitral: Accuracy 100%, Precision 1.0000, Recall 1.0000, Specificity 1.0000, F1-score 1.0000

These promising results across all three models can help cardiologists and cardiothoracic surgeons rapidly identify valve type and specific brand from plain chest radiographs. Knowing the exact prosthesis type is clinically crucial because mechanical valves require lifelong anticoagulation with INR monitoring, while bioprosthetic valves do not — directly impacting patient management and bleeding risk.

We are now fine-tuning the models with larger datasets. This work shows how state-of-the-art yet easily deployable AI can improve cardiac care.