Lumbar spine anatomical annotation based on magnetic resonance images using artificial intelligence
A system for automated comprehensive assessment of clinical lumbar MRIs includes a MRI standardization component that reads MRI data from raw lumbar MRI files, uses an artificial intelligence (AI) model to convert the raw MRI data into a standardized format. A core assessment component automatically generates MRI assessment results, including multi-tissue anatomical annotation, multi-pathology detection and multi-pathology progression prediction based on the structured MRI data package. The core assessment component contains a semantic segmentation module that utilizes a deep learning artificial intelligence (AI) model to generate an MRI assessment results that contains multi-tissue anatomical annotation, a pathology detection module to generate multi-pathology detection, and a pathology progression prediction module to generate multi-pathology progression prediction. A model optimization component archives clinical MRI data and MRI assessment results based on comments provided by a specialist, and periodically optimizes the AI deep learning model of the core assessment component.
1. A system for fully automated comprehensive assessment of clinical lumbar magnetic resonance images, comprising:
a magnetic resonance images (MRI) standardization component that reads MRI data from raw lumbar MRI files with different formats, and converts the raw MRI data into standardized MRI data, and organizes the standardized MRI data into a structured MRI data package, wherein the standardized MRI data is a 3D array with a shape of (H, W,N), where H, W represent a height and width of each MRI slice, and N is a slice number of a MRI sequence, pixel value range of a MRI image array is normalized to [0,1) and a data type of the MRI image array is float 32;
a core assessment component automatically generates MRI assessment results which include multi-tissue anatomical annotation, multi-pathology detection and multi-pathology progression prediction based on the structured MRI data package, said core assessment component contains a semantic segmentation module that utilizes an unsupervised artificial intelligence (AI) deep learning pipeline to generate a multi-tissue anatomical annotation, a pathology detection module utilizes ana second AI deep learning pipeline to generate multi-pathology detection, and a pathology progression prediction module utilizes ana third AI deep learning pipeline to generate multi-pathology progression prediction; and
a model optimization component that archives clinical MRI data and MRI assessment results based on comments provided by a specialist, and periodically optimizes the AI deep learning pipelines of the core assessment component.
2. The system of claim 1 wherein the pathology detection module comprises:
an intervertebral disc (IVD) region locating unit that receives the multi-tissue anatomical annotation and determine a position of each IVD region, respectively;
an IVD region measuring unit that receives the multi-tissue anatomical annotation and determine a size of each IVD region, respectively;
an IVD region extracting unit that receives the position and the size of each IVD region, respectively, as well as the structured MRI data package, and extracts a three dimensional (3D) IVD region array of each IVD from an MRI image array, each IVD region array contains a whole corresponding IVD, two adjacent vertebral endplates, and a spinal canal within a posterior region of the IVD;
an IVD region standardizing unit that receives a 3D IVD region array of each IVD from the MRI image array and converts the received 3D IVD region arrays to standardized IVD region arrays, the received 3D IVD region arrays have different sizes depending on an actual anatomical dimensions of IVD and the IVD region standardizing unit adopts an interpolation technology to convert the sizes of received 3D IVD region arrays to a standard size; and
a pathology detection unit receives the standardized IVD region arrays and uses an AI deep learning model for an assessment of pathologies and provides the multi-pathology detection, the multi-pathology detection includes a classification and grading results for multiple pathologies within each IVD region array and the multi-pathology detection is the final output of the pathology detection module.
3. The system of claim 2 wherein the interpolation technology is one or more of nearest interpolation, linear interpolation, and bilinear interpolation.
4. The system of claim 2 wherein the IVD region of each intervertebral disc is a cuboid region, which is centered at and is parallel to the corresponding intervertebral disc, and wherein the IVD region locating unit determines the position of each IVD region by identifying a location and a rotation of each intervertebral disc based on the multi-tissue anatomical annotation; and
wherein the IVD region measuring unit determines the size of each IVD region by measuring dimensions of spinal tissues based on the multi-tissue anatomical annotation.