CoLumbo

Processes DICOM-formatted spinal MRI scans through a deep learning CNN model trained on large annotated spine imaging datasets to automatically detect and spatially localize common pathologies (disc herniation, stenosis, spondylolisthesis, fractures). The system generates confidence scores per finding and flags high-confidence anomalies for radiologist review, reducing manual scan review time by filtering normal or low-risk studies. Architecture likely uses multi-slice 3D convolution with attention mechanisms to capture anatomical context across vertebral levels.

Integrates with hospital PACS systems via DICOM API or HL7 messaging to automatically retrieve spinal MRI studies, process them through the detection model, and generate structured preliminary reports that populate radiology information systems (RIS). The system likely uses a message queue (e.g., AMQP, Kafka) to handle asynchronous processing of high-volume studies and maintains audit logs for regulatory compliance. Reports are formatted as HL7 or FHIR-compliant structured data that radiologists can import, review, and electronically sign.

Provides a web or desktop interface where radiologists review AI-generated findings, adjust confidence thresholds, add clinical context, and electronically sign final reports. The system tracks radiologist edits and model predictions side-by-side, enabling feedback loops to retrain or fine-tune the model on institutional data. Implements role-based access control (radiologist, attending, administrator) and maintains immutable audit trails for regulatory compliance. Likely uses a collaborative annotation UI with keyboard shortcuts and voice dictation for efficient report finalization.

Generates per-finding confidence scores (0-1 scale) for multiple spinal pathologies (disc herniation, stenosis, spondylolisthesis, fractures, etc.) and aggregates them into a study-level risk stratification (normal, low-risk, moderate-risk, high-risk). The scoring likely uses Bayesian uncertainty quantification or ensemble methods (multiple model predictions) to estimate model confidence rather than raw softmax probabilities. High-risk studies are automatically prioritized for radiologist review, enabling triage-based workflow optimization.

Automatically identifies and localizes vertebral levels (C1-L5), intervertebral discs, spinal cord, and nerve roots in 3D space using semantic segmentation or keypoint detection networks. This enables spatial grounding of pathology findings (e.g., 'L4-L5 disc herniation' rather than generic 'disc herniation') and supports automated measurement of stenosis severity or disc height. Architecture likely uses U-Net or similar encoder-decoder networks with 3D convolutions to preserve volumetric context.

Compares current spinal MRI studies with prior imaging (weeks to years prior) to detect interval changes in pathology severity, new findings, or resolution of previously identified abnormalities. Uses image registration (rigid or deformable) to align current and prior studies in 3D space, then applies difference detection algorithms to highlight regions of change. Enables longitudinal tracking of degenerative disc disease progression, post-surgical healing, or treatment response.

Converts AI-generated findings and radiologist-validated annotations into standardized structured data formats (HL7 FHIR, DICOM SR, or proprietary JSON) that can be ingested by downstream clinical systems (EHR, surgical planning software, research databases). Uses schema-based extraction with predefined ontologies for spinal pathologies, severity grades, and anatomical locations. Enables automated population of structured fields in EHR systems and supports clinical decision support rules (e.g., 'if severe stenosis at L4-L5, flag for neurosurgery consultation').