Liver tumor detection based on objects as points.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate liver tumor detection in computed tomography (CT) is challenging due to variations in size, location, and shape.
• Current two-stage detection methods require bounding box annotations and can be slow and inefficient.
• Existing methods struggle with precise tumor center point prediction and can generate redundant bounding boxes.

Purpose of the Study:

• To develop a novel single-stage detector for simultaneous, accurate, and efficient detection of multiple liver tumors in CT images.
• To improve upon existing methods by enhancing speed, efficiency, and prediction accuracy, particularly for tumor center points.
• To create a cost-effective solution suitable for widespread clinical adoption, including in resource-limited settings.

Main Methods:

• Proposed a single-stage detector model for liver tumor detection.
• Implemented a multi-channel approach for CT image analysis to capture continuity information.
• Incorporated a bounding box attention mechanism to refine tumor center point prediction and reduce false positives.
• Utilized a Squeeze-and-Excitation attention module for effective information integration across channels.

Main Results:

• The proposed single-stage detector achieved a mean average precision (mAP) of 0.476 on the Decathlon dataset.
• Demonstrated superior speed and efficiency compared to existing two-stage detection methods.
• Successfully improved the accuracy of tumor center point prediction and decreased redundant bounding boxes.

Conclusions:

• The developed single-stage detector offers a faster, more efficient, and accurate approach for liver tumor detection in CT scans.
• The method's ability to accurately predict tumor center points can significantly aid physicians in rapid diagnostic verification.
• The model's high performance with low computational cost makes it suitable for clinical practice and resource-poor areas.