Automatic Identification of Anatomical Locations for Bone Abnormalities in CT Imaging: A Multiplanar YOLOv5 Detection...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning aids in detecting bone abnormalities like tumors and osteolytic lesions.
• Accurate anatomical localization is crucial for diagnosing bone abnormalities.

Purpose of the Study:

• To develop and evaluate an automatic bone detection and classification method for CT scans.
• To determine the anatomical location of bone abnormalities using a multiplanar deep learning approach.

Main Methods:

• Utilized three 2D YOLOv5l models for bone detection and classification across axial, coronal, and sagittal planes.
• Benchmarked the proposed method against nnU-Net using TotalSegmentator and RibFrac datasets.
• Evaluated performance in identifying osteolytic lesions and rib fractures.

Main Results:

• The proposed method outperformed nnU-Net in identifying non-vertebrae/rib lesions.
• nnU-Net excelled in precise vertebra-level and rib-fracture localization.
• The multiplanar method showed high accuracy in predicting ranges of rib or vertebra levels, surpassing nnU-Net.

Conclusions:

• The multiplanar bone detection approach is a competitive alternative to 3D segmentation for bone abnormality identification in CT scans.
• This method provides reliable localization and anatomical information for bone abnormalities, aiding reporting and research.