we tried to reconstruct a 3D volume out of a freehand ultrasound measurement.
We captured a video using a frame grabber and imported as video camera stream.
After recording we converted the video / image stream into a processed ultrasound clip.
We checked the option convert to sweep.
When we try to run the alogrithm sweep compounding this error occurs:
Cannot compound into degenerate volume of size 1 x 334 x 213.
How can we select the frames, because it appears only one image is used.
We tried the example from you GitHub repository (public-ultrasound-demos/Compounding) and recognized the meta data from your sweep (US-Forearm-Sweep-h264.imf) has elements which are missing in sweep file.
The issue is not the missing data components, but the missing tracking information. The Sweep Compounding algorithm requires pose information for each frame, or a tracking sequence (e.g., from an optical or EM tracking system, or from a robot). The error message indicates that this is not the case because all frames appear to be at the same position.
we don’t have any tracking informations, because we want to reconstruct a 3D volume just based on a freehand ultrasound scan.
Do we have to estimate the tracking informations and add them to the Sweep before using the Sweep Compounding algorithm, or is the workflow for freehand ultrasound reconstruction different?
We already used an optical flow algorithm to estimate the in-plane motion. However we are still struggling with the out-of-plane motion for the remaining axis. In a paper ImFusion proposes a method Speckle Decorrelation for this specific task.
Is this algorithm also implemented in the ImFusion Suite? In my understanding we need this algorithm for the full sensorless freehand ultrasound 3d reconstruction workflow.
Well, yes, and no. The algorithm is technically part of the SDK, but we’re not shipping any ML model because such models need to be fine-tuned for the hardware (US system, transducer), and imaging settings (in particular speckle filter) you are using, so our internal model is likely to anyway perform poorly on your data.
However, we’re currently preparing a submission to the TUS-REC challenge 2025 and will make the pipeline and the models public.
