3D Label Mask Synchronization With Deterministic GPU Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for synchronizing high-complexity virtual environment states in collaborative sessions, such as medical image processing, fail to ensure accurate replication of 3D label masks across different GPUs due to deviations in floating-point operations, leading to inconsistent virtual environments.
Innovation Solution
A method using bitwise operations on GPUs to compute ellipsoids proportional to voxel shapes, ensuring deterministic replication of 3D label masks by approximating inverse divisions, thereby maintaining consistent label mask replication across diverse hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard floating-point operations are used on different GPUs to compute 3D label masks, then computation speed is maintained, but replication accuracy deteriorates due to deviations in floating-point operations across different hardware
Solution Approach 1:
The patent changes the computational parameters from standard floating-point operations to bitwise operations. Specifically, it transforms the division operation into a multiplication by an approximated inverse value, and implements this using bitwise AND operations on the binary representation of floating-point numbers. This parameter change maintains computational speed while ensuring deterministic results across different GPU hardware, thereby improving replication accuracy without significantly increasing device complexity.
Solution Approach 2:
The patent substitutes the mechanical floating-point division operation with a bitwise operation system. Instead of using the GPU's floating-point unit to perform division, the system represents floating-point numbers in binary form and uses bitwise AND operations to compute the product of a number and an approximated inverse. This substitution eliminates hardware-dependent floating-point behavior while maintaining computational efficiency, resolving the contradiction between accuracy and complexity.
2Reliability
If high-complexity virtual environment states are synchronized in real-time, then collaboration quality is improved, but network bandwidth consumption increases due to large data transmission requirements
Solution Approach 1:
The patent extracts and transmits only the essential segmentation input data (brush position, radius, and label identifier) rather than transmitting entire 3D label masks. By separating the critical segmentation parameters from the full volumetric data, the system enables receiving devices to locally compute label mask updates using deterministic bitwise operations. This extraction approach maintains collaboration consistency while dramatically reducing network bandwidth consumption, as only small parameter sets are transmitted instead of large volumetric datasets.
3Measurement precision
If deterministic computation is implemented using bitwise operations, then replication accuracy is improved, but computation speed may deteriorate due to additional processing steps
Solution Approach 1:
The patent performs preliminary computation of the inverse value once, storing it for repeated use in multiple multiplication operations. By pre-computing the inverse of the divided value and reusing it across different voxel calculations, the system eliminates the need for repeated division operations. This preliminary action ensures deterministic results through bitwise operations while maintaining computation speed, as the expensive inverse computation is performed only once rather than for each voxel.
Data Source
AI summary
Computing device, method and computer program product for replicating a three-dimensional label mask for a three-dimensional volume dataset in a network of computing devices. A first device receives a series of segmentation input data from a further computing device of the network describing an editing of the three-dimensional label mask on a second computing device. The segmentation input data comprises current position data, size data and a label identifier associated with a sphere-shaped segmentation object used by the second computing device for interactively editing the label mask. A replicator module replicates the edited label mask of the second computing device by computing a respective ellipsoid and determining voxels of the three-dimensional label mask having a position inside the segmentation object, and updating, in the replicated label mask, the determined voxels in accordance with the label identifier of the segmentation input data.