Ensembled YOLO for multiorgan detection in chest x-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computer Vision

Background:

• Chest radiographs (CXRs) are crucial for diagnosing thoracic pathologies.
• AI/ML applications in CXR analysis require precise anatomical structure detection.
• Current object detection models like YOLO excel at single organ detection but lack multi-organ capabilities.

Purpose of the Study:

• To develop and evaluate a multi-organ detection technique for simultaneous lung and heart identification in CXRs.
• To compare the performance of recent YOLO versions (YOLOv8 and YOLOv9) and their sub-variants for this task.
• To assess the generalizability of the proposed models using external datasets.

Main Methods:

• Implementation of multi-organ detection using two recent YOLO versions and their sub-variants.
• Training and internal validation on the JSRT CXR dataset.
• External evaluation on the Montgomery CXR and RSNA CXR datasets.
• Ensemble approaches were explored to improve detection performance.

Main Results:

• YOLOv9 models significantly outperformed YOLOv8 variants in simultaneously detecting lung and heart regions.
• The proposed multi-organ detection technique demonstrated effectiveness on both internal and external datasets.
• Ensemble methods further improved the detection accuracy.

Conclusions:

• YOLOv9 offers a robust solution for multi-organ detection in CXR images.
• The developed technique advances AI-driven medical image analysis for improved diagnostic capabilities.
• Simultaneous multi-organ detection in CXRs is feasible and beneficial for clinical applications.