Self-Attention Image Model for Colorectal CT Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for colorectal cancer (CRC) detection, such as MRI and computed tomography colonography (CTC), are limited by time-consuming bowel preparation and potential adverse reactions, hindering widespread use, while CT scans without preparation face challenges in accurately segmenting the colorectum and CRC due to organ crowding and small tumor size.
Innovation Solution
A deep colorectal coordinate conversion (DeepCRC) method is employed, using a one-dimensional coordinate system based on the colon and rectum centerline to improve segmentation accuracy by incorporating an auxiliary voxel-level regression task and a self-attention mechanism layer to model global context in image processing models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If CT scan without bowel preparation is used, then examination time and patient comfort are improved, but segmentation accuracy of colorectum and tumor deteriorates due to organ crowding and small tumor size
Solution Approach 1:
The patent introduces a one-dimensional coordinate system based on the centerline of the colorectum to transform the three-dimensional segmentation problem into a one-dimensional coordinate regression problem. This dimensional reduction allows the model to accurately locate tumors along the colorectal tract without requiring complex three-dimensional segmentation, thereby maintaining high accuracy while avoiding the need for bowel preparation.
Solution Approach 2:
The patent employs an auxiliary coordinate map as an intermediary representation between the input CT image and the final segmentation output. This coordinate map serves as a mediator that guides the segmentation process by providing explicit spatial constraints along the colorectal centerline, enabling accurate tumor localization even in crowded anatomical regions without bowel preparation.
2Measurement precision
If self-attention mechanism layer is added to image processing model, then global context modeling capability is improved, but model complexity increases
Solution Approach 1:
The patent divides the image processing model into distinct functional modules: an encoder for feature extraction, a self-attention mechanism layer for global context modeling, and a decoder for segmentation output. This modular segmentation allows the complex self-attention mechanism to be integrated systematically, improving global context understanding while managing model complexity through structured organization.
Solution Approach 2:
The self-attention mechanism operates in the feature space dimension rather than directly in spatial dimensions, transforming the way global context is modeled. By attending to relationships between features across the entire image, the model achieves comprehensive global context understanding without proportionally increasing spatial complexity.
Data Source
AI summary
A method, an apparatus, and a non-transitory computer readable medium for training an image processing model are provided. The method includes: acquiring a sample image comprising a target object to determine an object segmentation image of the target object in the sample image; constructing an object coordinate map corresponding to the object segmentation image according to the object segmentation image; and training an image processing model comprising a self-attention mechanism layer according to the sample image, the object segmentation image, and the object coordinate map.