MedYOLO: A Medical Image Object Detection Framework.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Computer Vision for Healthcare

Background:

• Convolutional Neural Networks (CNNs) are standard for medical image segmentation but require extensive expert annotation.
• Object detection models offer reduced annotation effort for tasks not needing voxel-level precision.
• Limited general-purpose 3-D object detection frameworks exist for medical imaging.

Purpose of the Study:

• Introduce MedYOLO, a 3-D object detection framework for medical imaging.
• Evaluate MedYOLO's performance across diverse medical datasets.
• Demonstrate the potential of object detection to reduce annotation burden in AI for radiology.

Main Methods:

• Developed MedYOLO, a 3-D object detection framework based on the YOLO (You Only Look Once) one-shot detection model.
• Tested MedYOLO on four distinct datasets: BRaTS, LIDC, abdominal CT, and ECG-gated heart CT.
• Assessed performance using mean average precision (mAP) at an intersection over union (IoU) threshold of 0.5.

Main Results:

• MedYOLO achieved high performance across multiple datasets without hyperparameter tuning.
• Achieved mAP@0.5 scores of 0.861 (BRaTS), 0.715 (abdominal CT), and 0.995 (heart CT).
• The model struggled with certain structures, failing to converge on the LIDC dataset (mAP@0.5 of 0.0).

Conclusions:

• MedYOLO is a promising 3-D object detection framework for medical imaging, offering efficiency and high performance.
• The framework demonstrates potential for reducing annotation effort compared to traditional segmentation models.
• Further development is needed to address challenges with specific datasets and structures, like LIDC.