Medical object detector jointly driven by knowledge and data.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Existing object detection algorithms struggle with ambiguous features in medical images, leading to misclassification and mislocalization.
• Reliance solely on data-driven features limits accuracy when visual cues are subtle or absent.

Purpose of the Study:

• To develop a novel medical object detection algorithm (ODKD) that integrates expert medical semantic knowledge with data-driven approaches.
• To enhance the robustness and accuracy of object detection in medical imaging, particularly in challenging cases.

Main Methods:

• Proposed a hybrid model (ODKD) comprising a base object detector and a knowledge-data fusion module.
• Utilized a graph structure to represent medical semantic knowledge and a graph convolution network to fuse it with data-driven results.
• Generated category adjustment coefficients to refine the initial predictions from the base detector.

Main Results:

• The ODKD model effectively corrected erroneous detections by leveraging external medical semantic knowledge.
• Demonstrated superior performance compared to established models like Faster RCNN, YOLOv5, and YOLOv7 on the Camus, Synapse, and AMOS datasets.
• Validated the significant impact of incorporating professional medical knowledge in object detection tasks.

Conclusions:

• Integrating medical semantic knowledge significantly improves the accuracy and reliability of object detection in medical imaging.
• The proposed ODKD algorithm offers a promising solution for enhancing diagnostic accuracy in scenarios with subtle or ambiguous image features.
• ODKD outperforms traditional methods, highlighting the value of hybrid knowledge-and-data-driven approaches in medical AI.