Problem

Computed tomography (CT) is commonly used for its widespread availability, rapid acquisition times, and depiction of fine bony detail, but it struggles to differentiate spinal cord tissue from cerebrospinal fluid (CSF) due to their similar mass densities. Magnetic resonance imaging (MRI) offers superior intraspinal soft tissue and fluid depiction but is often unavailable or contraindicated. Myelography provides high-resolution imaging but is invasive and time-consuming. Therefore, there is a clinical need for alternative CT-based approaches to visualize intraspinal soft tissues more accurately

Solution

Researchers at OSU have developed a method to improve visualization of soft tissue and water pixels on CT. The invention includes a machine learning-based method for applying convolutional neural networks trained on small 3D patches of CT data to perform a voxel classification task and produce a 4-class (CSF, spinal cord tissue, fat, bone) segmentation map from CT images. In the current implementation, the model takes 2 CT volumes, acquired at low (100 kVp) and high (140 kVp) energies, and generates output images designed to accentuate the differences between intraspinal cerebrospinal fluid and intraspinal soft tissue (spinal cord and nerve roots). An ensemble model incorporating models trained on different patch sizes is used to perform the final 4-class classification task. The tool has high sensitivity and specificity for distinguishing between spinal cord tissue and CSF (99–100%), and bone and fat voxels (99.6-100%)

Applications

• Clinical interpretation of spinal CT data when MRI is contraindicated or unavailable
• Non-invasive alternative to traditional myelography
• Preoperative planning of spinal surgeries
• Radiology workflow to improve soft tissue visualization without contrast agents

Advantages

• Virtual Myelography is non-invasive, does not require contrast injection and is compatible with routine lumbar spine CT data
• The machine learning model applied to dual-energy CT data can accurately differentiate among various tissue types based on the center voxel and neighboring voxels, beyond the capability of radiologists and standard thresholding approaches
• While the high voxel classification sensitivities and specificities for soft tissue and CSF represent most of the novelty of the invention, the performance in classification of bone and fat voxels also represents improvement from currently available methods

Seeking opportunities for co-development, out-licensing or new venture formation