Two-Stage CNN Framework for Thoracic Organ Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for automatic thoracic organ segmentation in CT images are time-consuming, prone to human error and variability, and require large, diverse training datasets, with challenges in handling incomplete data and systematic differences across clinical sites.
Innovation Solution
A two-stage deep learning framework using convolutional neural networks (CNNs) for thoracic organ segmentation, which processes 3D CT images to uniform spatial resolution, adapts to incomplete training data, and improves with new data, incorporating data augmentation and transfer learning to address variability and performance issues across sites.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual contouring is used for organ segmentation, then segmentation can be performed with existing methods, but it takes 1-2 hours per patient and suffers from inter- and intra-observer variability
Solution Approach 1:
The patent replaces the manual mechanical contouring process with an automated deep learning system. The CNN-based framework automatically segments thoracic organs from CT images, eliminating the need for manual contouring while achieving both high speed and consistent results. The system processes images through multiple stages including preprocessing, segmentation network inference, and post-processing to generate final organ masks.
Solution Approach 2:
The deep learning system performs segmentation autonomously without requiring manual intervention. The framework independently processes CT images, applies data augmentation, and generates segmentation results automatically. The system serves itself by handling the complete segmentation workflow from input images to output organ masks without human contouring effort.
2Extent of automation
If traditional deep learning frameworks are used, then automation can be achieved, but they heavily rely on large and diverse training datasets which are difficult to collect incrementally
Solution Approach 1:
The patent implements a dynamic training framework that adapts to newly available data. The system allows incremental updates to the training dataset and retraining of the segmentation network. This dynamic approach enables the model to improve its performance over time as more data becomes available, rather than requiring a complete static dataset upfront.
Solution Approach 2:
The framework performs preliminary data augmentation and preprocessing steps before training to prepare the data efficiently. By pre-processing images with various augmentations (rotation, flipping, scaling) and organizing them into structured formats, the system prepares the training data in advance, making the subsequent model training more efficient and adaptable to new data arrivals.
3Ease of manufacture
If deep learning models are trained on incomplete training data, then training can proceed with available data, but the models may not achieve optimal performance
Solution Approach 1:
The patent applies partial training approaches where the model is trained on the subset of available data rather than waiting for complete datasets. The framework performs data augmentation to maximize the utility of limited training samples and uses transfer learning to compensate for incomplete data. This allows the system to achieve reasonable performance with partial data while maintaining the option to improve further as more data becomes available.
4Adaptability or versatility
If data from multiple clinical sites are collected, then dataset diversity increases, but systematic differences between sites may affect model performance
Solution Approach 1:
The patent applies site-specific preprocessing and normalization techniques to handle systematic differences between clinical sites. The framework identifies and corrects site-specific variations in imaging protocols, scanner types, and patient populations through targeted preprocessing steps. This local quality adjustment ensures that each site's data is appropriately normalized before being fed into the segmentation model, maintaining consistent performance across diverse clinical environments.
Data Source
AI summary
The present disclosure relates to a method and apparatus for automatic thoracic organ segmentation. The method includes: receiving three-dimensional (3D) images obtained by a computed tomography (CT) system; processing the 3D images to have the same spatial resolution and matrix size; building a two-stage deep learning framework using convolutional neural networks (CNNs) for organ segmentation; adapting the deep learning framework to be compatible with incomplete training data; improving the CNNs upon arrival of new training data; post-processing the output from the deep learning framework to obtain final organ segmentation.