Deep convolutional neural network-based segmentation and classification of difficult to define metastatic spinal...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Spinal metastatic lesions, both lytic and sclerotic, are challenging to segment and classify using traditional methods due to ill-defined boundaries.
• Classical texture and shape analysis methods struggle with feature extraction from complex spinal CT images.

Purpose of the Study:

• To develop an automated method for segmenting and classifying lytic and sclerotic metastatic spinal lesions.
• To improve the detection and characterization of difficult-to-define lesions in 3D Computed Tomography (CT) scans.

Main Methods:

• Utilized a deep Convolutional Neural Network (CNN) for automatic feature extraction.
• Implemented patient-data and scan-protocol-dependent pre-processing steps for adaptability to various CT scans.
• Employed medial axis transform (MAT) post-processing and Random Forest (RF) based meta-analysis for shape simplification and classification.
• Applied the method to whole-spine CT scans (cervical, thoracic, lumbar).

Main Results:

• The proposed CNN-based method successfully segmented and classified metastatic lesions.
• The approach demonstrated usability on whole-spine CT datasets, outperforming methods limited to thoracolumbar segments.
• Validation involved a dataset annotated by two independent radiologists and comparison with existing methods.

Conclusions:

• The developed deep learning framework effectively addresses the segmentation and classification of challenging spinal metastatic lesions.
• This method offers a robust solution for analyzing lesions across the entire spine, advancing spinal lesion longitudinal studies.