MTASololu | Devpost

Landing Page
Logo

Inspiration

Annotating medical imaging is critical for training, analysis, and collaboration, but static annotations lose alignment when anatomy or cameras move. We wanted to create annotations that stay aligned with anatomical features, making video review more accurate and reliable for clinicians, researchers and surgeons.

What it does

MTASololu processes uploaded medical videos and generates motion-tracked annotations that stay fixed relative to the anatomy. As structures move throughout the video, the annotations follow, enabling precise reference points for analysis, training, or collaborative review.

How we built it

We extract video frames and use OpenCV to detect key features. Motion is tracked via optical flow, and a lightweight neural network refines annotation alignment across frames. The system outputs the video with all annotations correctly positioned throughout.

Challenges we ran into

Processing noisy or low-contrast medical videos
Maintaining high accuracy while keeping processing time reasonable
Supporting multiple imaging modalities (ultrasound, echocardiography, laparoscopy)
Ensuring annotations remain consistent even with camera motion

Accomplishments that we're proud of

Developed a hybrid CV + ML pipeline that works across modalities
Created annotations that stay aligned frame-to-frame without manual adjustment
Built a scalable workflow for batch processing uploaded videos

What we learned

Combining classical computer vision and machine learning can create robust tracking
Preprocessing and video quality greatly affect annotation accuracy
Careful system design is required to generalize across different medical imaging types

What's next for MTASololu

Support real-time streaming for live procedures
Integrate with HoloXR for collaborative annotation and review
Expand training datasets to improve neural network tracking accuracy
Add interactive editing tools for clinicians to adjust annotations when needed