3D Bounding Box Detection for GPU Memory Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based 3D organ at risk (OAR) segmentation is hindered by limited GPU memory resources, leading to inefficiencies such as high frequency information loss and increased processing time when using downsampling or manual cropping techniques.
Innovation Solution
The method involves detecting a 3D bounding box using 2D maximum intensity projection (MIP) images from axial and sagittal views, reducing the input volume to the region of interest, thereby conserving GPU memory and processing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If downsampling is used as preprocessing, then GPU memory consumption is reduced, but high frequency information is lost and segmentation accuracy decreases
Solution Approach 1:
The patent segments the 3D volume data by detecting a bounding box around the region of interest (ROI) containing the organ at risk. This divides the large volume into a smaller sub-volume that fits within GPU memory, allowing full-resolution processing of the relevant region without downsampling the entire dataset.
Solution Approach 2:
The patent extracts only the necessary portion of the 3D volume data by identifying and isolating the bounding box that contains the organ at risk. This extraction removes unnecessary data from the processing pipeline, reducing memory requirements while preserving all high-frequency information within the ROI.
2Quantity of substance
If manual cropping is used to reduce input size, then GPU memory consumption is reduced, but human interaction is introduced and efficiency decreases
Solution Approach 1:
The system performs automatic bounding box detection using deep learning models that identify the organ at risk and its surrounding region. This self-service approach eliminates the need for manual cropping by operators, automating the ROI identification process and improving processing efficiency while adapting to different anatomical variations.
Solution Approach 2:
The patent changes the parameter of ROI identification from manual coordinate specification to automated detection based on image content analysis. The deep learning model automatically determines the bounding box parameters (position, size, orientation) by analyzing the 3D volume data, eliminating manual intervention while maintaining adaptability to different cases.
3Manufacturing precision
If sliding window ensemble methods are used, then segmentation accuracy is improved, but processing time increases significantly
Solution Approach 1:
The patent performs preliminary action by detecting the bounding box of the organ at risk before the segmentation process. This pre-processing step identifies and isolates the ROI, allowing the subsequent segmentation network to focus computational resources on a smaller, relevant region. This eliminates the need for sliding window ensemble methods across the entire volume, significantly reducing inference time while maintaining accuracy.
Data Source
AI summary
A method and device for generating a three dimensional (3D) bounding box of a region of interest (ROI) of a patient include receiving a two dimensional (2D) maximum intensity projection (MIP) image that is an axial view of the patient and a 2D MIP image that is a sagittal view of the patient. A first 2D bounding box of the ROI of the patient and a second 2D bounding box of the ROI of the patient are detected using the 2D MIP images. A 3D MIP image of the patient is received, and the 3D bounding box of the ROI of the patient is generated using the 3D MIP image, the first 2D bounding box, and the second 2D bounding box. The 3D MIP image including the 3D bounding box is provided.