Lesion Detection Using 2D-3D Overlap Against Slice Interval Variation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting lesion regions using three-dimensional volume data are susceptible to decreased accuracy due to variations in slice intervals between training and inference, particularly when the slice interval during inference is larger than during training.
Innovation Solution
A method that combines lesion detection from both two-dimensional tomographic images and three-dimensional volume data by using machine learning models to identify overlapping regions, thereby reducing the impact of slice interval variations on detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If three-dimensional volume data is used for lesion detection, then detection accuracy is improved, but reliability deteriorates when slice interval is larger than training interval
Solution Approach 1:
The patent combines 2D lesion detection results and 3D volume data detection results into a final detection output. The 2D model processes individual tomographic images while the 3D model processes volume data, and their results are merged through a combination unit that generates the final lesion detection output, thereby leveraging the strengths of both approaches to compensate for slice interval variations.
Solution Approach 2:
The patent introduces a slice interval adjustment unit that adjusts the slice interval of tomographic images based on the detected lesion region. By dynamically changing the slice interval parameter according to the lesion characteristics, the system adapts to different imaging conditions and maintains detection accuracy even when the inference slice interval differs from the training interval.
2Shape
If 3D volume data is used, then three-dimensional shape feature capture is improved, but sensitivity to slice interval variations increases
Solution Approach 1:
The system merges 2D and 3D detection results to create a more robust detection output. The 2D model provides detailed information from individual slices while the 3D model captures volumetric shape features, and their combination allows the system to maintain shape accuracy while being less sensitive to slice interval variations through cross-validation of detection results.
Solution Approach 2:
The slice interval adjustment unit dynamically modifies the slice interval based on the lesion region detected by the 2D model. This dynamic adjustment allows the system to adapt to varying imaging conditions and maintain optimal detection performance regardless of the fixed slice interval used during data acquisition, thereby improving slice interval adaptability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A lesion detection method for a computer to execute a process includes detecting a first lesion region that indicates a certain lesion from each of a plurality of tomographic images obtained by imaging an inside of a human body, by using a first machine learning model that detects a lesion region from an input image that has image data of a two-dimensional space; detecting a second lesion region that indicates the certain lesion from three-dimensional volume data generated based on the plurality of tomographic images, by using a second machine learning model that detects a lesion region from input volume data that has image data of a three-dimensional space; and detecting a third lesion region that indicates the certain lesion from each of the plurality of tomographic images, based on an overlapping state between the first lesion region and the second lesion region.