Efficient organ localization using multi-label convolutional neural networks in thorax-abdomen CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate organ localization is crucial for medical image analysis tasks like segmentation and registration.
• Existing methods may lack efficiency or accuracy in localizing multiple structures simultaneously.

Purpose of the Study:

• To develop and evaluate an efficient method for simultaneous localization of multiple structures in 3D thorax-abdomen CT scans.
• To improve the accuracy and efficiency of organ localization using deep learning.

Main Methods:

• Utilized a single multi-label convolutional neural network (CNN) for each orthogonal view to predict multiple structure locations.
• Incorporated extra contextual slices as input to enhance network performance.
• Combined outputs from three CNNs to generate 3D bounding boxes for 11 structures of interest.

Main Results:

• Achieved an average wall distance of [Formula: see text] mm on a dataset of 1884 CT scans, outperforming previous studies.
• The method's performance was comparable to human observers ([Formula: see text] mm).
• Demonstrated improved performance by utilizing multiple slices for contextual information.

Conclusions:

• Proposed an efficient and accurate method for localizing multiple organs in medical images.
• The use of contextual slices significantly improves localization performance.
• The method is adaptable for localizing a greater number of organs.