Deep learning-based body part recognition algorithm for three-dimensional medical images.

Area of Science:

• Medical imaging analysis
• Deep learning in radiology
• Computational anatomy

Background:

• Automatic recognition of human body parts in 3D medical images is crucial for clinical applications.
• Prior methods often classified 2D slices independently, lacking contextual understanding of consecutive slices.

Purpose of the Study:

• To develop a deep learning method for automatic division of CT and MRI scans into five consecutive body parts: head, neck, chest, abdomen, and pelvis.
• To improve the accuracy of body part recognition in medical imaging.

Main Methods:

• A two-stage deep learning framework combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) for slice classification.
• Utilized GoogLeNet Inception v3 architecture for CNN feature extraction and LSTM for contextual information from consecutive slices.
• A postprocessing stage identified optimal boundaries for partitioning scans into body parts.

Main Results:

• Achieved high 2D slice classification accuracies: 97.3% for CT and 98.2% for MRI.
• Accurately divided whole scans with mean boundary errors of 8.9 mm for CT and 3.5 mm for MRI.
• Demonstrated superior performance compared to state-of-the-art methods.

Conclusions:

• The developed method significantly enhances slice classification accuracy over existing techniques.
• Accurately segments CT and MRI scans into consecutive body parts, aiding computer-aided diagnosis.
• Provides a valuable tool for medical image analysis and clinical workflows.