Building a 3D Spine Model from 2D Ultrasound
Challenge
The client was developing an AI solution for spinal ultrasound (USG) imaging to assist in clinical procedures. The procedures require access to the spinal cord, meninges, and cerebrospinal fluid (CSF) spaces. However, ultrasound is not the preferred modality for imaging bones, making it inherently challenging to identify and delineate bony structures like vertebrae and ribs.
Solution
Cogito Tech’s annotation experts collaborated closely with a certified radiologist to identify key bony landmarks based on clinical morphological criteria. Despite ultrasound's limitations, we successfully annotated vertebrae using advanced 3D interpolation tools and incorporated CT overlays to guide and validate complex anatomical features. This hybrid approach improved annotation precision and efficiency.
Outcome
We delivered high-quality, clinically verified annotations that enabled the client to provide the world’s first FDA-approved AI-powered 3D spinal, real-time, and radiation-free ultrasound solutions. This breakthrough contributed to safer instrument navigation during spinal procedures and streamlined the model training process. Our work boosted the client’s AI model to 99% accuracy and accelerated development timelines.
How Cogito Tech Started?
Cogito Tech started working on this project by deeply understanding the client’s objective to create an AI-based, FDA-approved, and radiation-free spinal ultrasound imaging system. Since there is a challenge in visualizing bony structures in ultrasound images, our team began collaborating with a certified radiologist to establish proper annotation guidelines. We started with a pilot group of images to fine-tune the annotation process, applying clinical morphological expertise to detect vertebrae reliably. After being standardized on frameworks, we proportioned the annotation using sophisticated 3D interpolation software and CT reference overlays, minding precision, standardization, and medical applicability throughout.
Steps We Followed
The success of the project was ensured with the collaboration of our experienced annotators, the guidance of a certified radiologist, and the following steps:-
- Medical Image Assessment
The dataset included multi-slice DICOM images from X-rays to convert the images into 3D. Cogito Tech’s medical data specialists evaluated the scan types, resolutions, and anatomical variability to define a consistent annotation protocol. - Custom Annotation Guidelines
We developed anatomy-specific labeling protocols with radiology advisors (e.g., organs, bones, soft tissue boundaries). Each guideline was aligned with clinical relevance and cross-slice requirements for 3D views. - Tool Integration & Training
Advanced annotation tools with support for DICOM, NIfTI, and volumetric view handling were deployed. In parallel, annotators worked on 2D-3D alignment, using reference slices and landmarks to ensure anatomical fidelity. - 2D Slice-by-Slice Labeling
Our team meticulously annotated slices per patient, marking organs, pathologies, and tissue boundaries. These labels formed the foundation for volumetric interpolation and 3D model training. - 3D Volume Verification
Annotations were validated within 3D reconstructed views to ensure inter-slice alignment. Any inconsistencies, such as shifts or shape deformations, were corrected through iterative refinements and AI-assisted QC (quality checks). - Automated & Expert QC
Multi-stage quality control is combined with algorithmic validation (checking slice overlap and boundary precision) and expert human review by medical imaging professionals to maintain clinical accuracy.
Results & Impact
We helped the client build the world’s first FDA-cleared, AI-powered 3D spinal ultrasound solution, offering real-time, radiation-free imaging for safer and more accurate clinical procedures.
The model substantially amplified accuracy with reduced error rates and improved performance in critical tasks like volumetric analysis of different parts of the spine, such as vertebrae (bones), soft tissues, nerves, and more. These improvements supported faster diagnoses and higher physician confidence.
Are you searching for the best image annotation partner for your next project? Let’s connect today!