Stratified Image Segmentation for Radiotherapy OARs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation methods for organs at risk (OARs) in head and neck cancer radiotherapy face challenges due to complex anatomical shapes, dense spatial distributions, large size variations, and low image contrast, leading to inefficiencies and high inter-practitioner variability in manual delineation, and difficulties in model optimization for automatic segmentation.
Innovation Solution
A stratified image segmentation method using a progressive holistically-nested neural network (P-HNN) framework that categorizes OARs into anchor-level, mid-level, and small-level objects, employing different neural network models for each category, with anchor-level objects serving as references for mid-level and small-level object segmentation, and incorporating tailored segmentation workflows and network architecture search for optimized performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual delineation is used for OAR segmentation, then segmentation can be performed with human expertise, but it is time consuming and suffers from large inter-practitioner variability
Solution Approach 1:
The patent replaces the manual mechanical delineation process with an automated neural network-based system. The P-HNN model automatically segments OARs from RTCT images, eliminating the need for manual tracing by radiation oncologists. This substitution maintains high segmentation accuracy while dramatically reducing the time required, addressing both the precision and time consumption issues of manual delineation.
2Productivity
If whole volume segmentation is performed, then all OARs are segmented in one process, but model optimization becomes difficult and performance is not tailored to challenging problems
Solution Approach 1:
The patent divides the OAR segmentation task into three hierarchical levels: anchor-level OARs (large, easy-to-locate organs), mid-level OARs (moderate size and complexity), and small-level OARs (small, difficult-to-detect organs). Each level is processed by a specialized neural network model optimized for its specific characteristics. This segmentation approach improves both productivity by enabling parallel processing and reliability by tailoring each model to its specific segmentation challenges.
3Measurement precision
If segmentation methods are designed for complex anatomical shapes and low contrast, then accurate segmentation can be achieved, but model optimization becomes challenging
Solution Approach 1:
The patent applies local quality by designing different neural network models with specific architectures and loss functions tailored to each OAR level's characteristics. Anchor-level models use standard architectures, mid-level models incorporate spatial attention mechanisms, and small-level models use feature pyramid networks. This localized optimization for each segment's specific challenges (size, contrast, shape complexity) achieves high accuracy without requiring a single overly complex universal model.
Data Source
AI summary
A method and device for stratified image segmentation are provided. The method includes: obtaining a three-dimensional (3D) image data set representative of a region comprising at least three levels of objects; generating a first segmentation result indicating boundaries of anchor-level objects in the region based on a first neural network (NN) model corresponding to the anchor-level objects; generating a second segmentation result indicating boundaries of mid-level objects in the region based on the first segmentation result and a second NN model corresponding to the mid-level objects; and generating a third segmentation result indicating small-level objects in the region based on the first segmentation result, a third NN model corresponding to the small-level objects, and cropped regions corresponding to the small-level objects.