Two-Stage DCNN for OAR Auto-Contouring in Medical Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for delineating organs at risk (OARs) in radiation therapy are time-consuming, error-prone, and unable to accurately process whole-volume 3D medical images, leading to false positives and inadequate anatomical variation handling.
Innovation Solution
A two-stage deep learning framework comprising an OAR Detection Network and an OAR Segmentation Network, utilizing 3D convolution and shared feature extractors, which operates on commodity GPUs, reducing memory costs and enabling efficient segmentation of OARs in the head, neck, and abdomen regions without relying on image templates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current deep learning-based OAR delineation methods (U-Net, AnatomyNet) are used to segment OARs on a slice-by-slice basis, then the system can process images without requiring whole-volume processing, but this produces false positives and cannot constrain rough locations and sizes of OARs
Solution Approach 1:
The patent divides the OAR delineation task into two distinct stages: a detection stage that identifies candidate OAR locations and sizes, and a segmentation stage that refines the boundaries. This segmentation of the processing pipeline allows the system to first constrain the search space (reducing false positives) and then perform detailed segmentation, resolving the contradiction between reliability and complexity
Solution Approach 2:
The detection network performs preliminary action by identifying and constraining the rough locations and sizes of OARs before the segmentation stage. This preliminary constraint on search space prevents false positives from occurring in the segmentation phase, thereby improving reliability without requiring overly complex whole-volume processing
2Measurement precision
If manual delineation by human experts is performed on 3D CT images, then accurate OAR boundaries can be obtained, but the process is time-consuming and error-prone
Solution Approach 1:
The system enables self-service by having the detection network automatically identify candidate OAR locations, sizes, and classifications without human intervention. The segmentation network then refines these automatically-generated candidates, allowing the system to serve itself in the initial delineation tasks while maintaining high accuracy and reducing time loss
3Adaptability or versatility
If image templates are used in current deep learning methods, then processing can be simplified, but the methods cannot adequately account for anatomical variations such as tumor growth
Solution Approach 1:
The patent implements dynamics by making the detection network learn anatomical variations dynamically from training data rather than relying on fixed templates. The network adapts to different anatomical configurations, tumor growth patterns, and organ shapes by learning from diverse examples, enabling it to handle anatomical variations while maintaining reasonable model complexity through efficient architecture design
4Manufacturing precision
If whole-volume 3D CT images are processed to capture accurate OAR shapes and sizes, then segmentation quality improves, but computational resources and GPU memory requirements increase significantly
Solution Approach 1:
The patent extracts and processes only the relevant candidate regions identified by the detection network, rather than processing the entire whole-volume 3D CT image. By taking out and focusing computational resources on constrained candidate locations and sizes, the system maintains high segmentation quality while significantly reducing GPU memory requirements and computational energy consumption
Data Source
AI summary
A system and methods for automatically delineating OARs in whole-volume medical images through a two-stage DCNN model, the DCNN model comprising an OAR detection network and an OAR segmentation network, and the method comprising the steps of: inputting the whole-volume medical images to the OAR detection network; extracting image features through a sequence of downsampling blocks; generating a final detection feature map via upsampling and concatenation; detecting at least one OAR candidate by branching the final detection feature map, wherein the at least one OAR candidate is defined by a predicted bounding box with a class label; inputting the predicted bounding box and corresponding class label to the OAR segmentation network; cropping the final detection feature map and a downsampling block in the OAR detection network according to the predicted bounding box; concatenating the cropped feature maps and generating a predicted binary mask delineating OARs according to the class label.