RPI presents a novel, DCL-Net system, and method based on a machine learning/DL technique to reconstruct 3D volumes from a series of 2D images. The developed technology does not require the use of sensor tracking hardware/devices (e.g., robotic arms, touchless position/pose trackers) to operate. The technology acquires multiple (two or more) consecutive, 2D US-generated image frames using a handheld imaging sensor or a sensor mounted on a motion restricted device used as inputs for estimating a US probe trajectory and tracking its motion. The 3D volume reconstruction technique is a type of freehand imaging technology and uses 3D convolutions over a TRUS video segment to extract key features and uses an embedded self-attention module to focus the network on speckle/visual-rich areas to provide a better estimate/prediction of spatial movement and geometry. The technology focuses on evaluating TRUS data/images for prostate cancer biopsy/diagnosis, though other applications are relevant.
The developed DCL-Net 3D volume reconstruction technology based on multiple, consecutive 2D US images and a generated video results in the prediction of more vivid, richer imaging features/geometries and fewer measurement errors than other techniques used for 3D volume reconstruction (e.g., Linear Motion, Decorrelation, 2D CNN, 3D CNN). Advantages of the technology include:
• The developed DCL-Net as used in 3D volume reconstruction operations realizes a better estimate/prediction of features identified in US images providing medical personnel a more efficient diagnostic tool for treating prostate cancer and other diseases.
• The developed 3D volume reconstruction technique eliminates the requirement to use a tracking device, allowing clinicians to obtain improved images by navigating the probe with fewer constraints.
• The developed DCL-Net 3D volume reconstruction technique can realize manufacturing cost efficiencies by eliminating the cost of sensor tracking hardware/devices (e.g., robotic arms, touchless position/pose trackers).
• Positioning errors from generated video segments are lower with the DCL-Net technology than commonly used 3D volume reconstruction methodologies.