Computerized Radiographic Imaging with Automated Projection Annotations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in obtaining sufficient, good quality, and diverse training data for data-driven techniques in medical radiological imaging, particularly in dental radiography, as creating large amounts of data pairs for supervised training and annotation is time-consuming and expensive.
Innovation Solution
Leverage existing radiological volumes from different patients, transforming them into images of predetermined modalities, and automatically providing annotation data for these images to enhance training data quality and robustness, including non-ideal behaviors and artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is performed to create training data pairs, then annotation quality can be ensured, but time consumption and cost increase significantly
Solution Approach 1:
The patent creates synthetic training data by copying and transforming existing annotated 3D radiological volumes. Multiple 2D projections are generated from each 3D volume, and annotations are automatically copied and transformed to match these projections. This eliminates manual annotation while preserving data quality through automated transformation of existing accurate annotations.
Solution Approach 2:
The patent performs preliminary annotation on 3D radiological volumes before generating 2D projections. By annotating the 3D volume first and then automatically transforming these annotations to corresponding 2D projections, the system prepares all necessary training data in advance without requiring subsequent manual annotation of each 2D image.
2Reliability
If diverse training data with various artifacts and distortions is created, then data-driven logic robustness improves, but data generation complexity increases
Solution Approach 1:
The patent changes parameters of existing radiological volumes by applying various transformations including rotation, scaling, flipping, and adding synthetic artifacts. These parameter changes create diverse training data with different orientations, sizes, and artifact types, improving model robustness while using automated processes rather than complex manual interventions.
Solution Approach 2:
The system performs self-service by automatically generating diverse training data with artifacts and distortions through programmed transformations. Instead of requiring complex manual creation of each variant, the system autonomously applies transformations and generates the necessary data diversity through automated processing pipelines.
3Quantity of substance
If existing radiological volumes are transformed to multiple image modalities, then training data quantity increases, but processing time increases
Solution Approach 1:
The patent performs preliminary transformation of a single 3D radiological volume into multiple 2D projection views and modalities before training. By pre-processing and transforming the volume once into all required formats (panoramic, bitewing, cephalometric, etc.), the system generates abundant training data in advance, reducing the need for repeated processing during model training.
Solution Approach 2:
The patent segments the 3D radiological volume into different projection views and modalities. Each 2D image is a segmented representation of the 3D volume from specific angles and perspectives. This segmentation approach efficiently generates multiple training samples from a single source volume by dividing it into distinct 2D projections.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present teachings relate to a method for improving a medical radiological process, comprising: providing a three-dimensional radiological volume; providing volume annotation data; wherein the volume annotation data comprise a plurality of annotations, transforming the radiological volume to a plurality of radiological projections; automatically providing, using the volume annotation data, projection annotation data for at least some of the radiological projections; reconstructing, from the plurality of radiological projections, a simulated radiological image; generating, using the projection annotation data, image annotation data. The present teachings also relate to systems, software products, storage media and uses.