Parametrized X-ray Anomaly Detection via GAN Topogram Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning algorithms require large amounts of X-ray images for training, but acquiring these images is challenging due to radiation exposure, and existing data augmentation methods focus on image alterations rather than patient-level augmentation, leading to insufficient variance in training datasets.
Innovation Solution
A system that generates and augments training data using parametrized X-ray images by creating synthetic patients through machine-learned networks, specifically generative adversarial networks (GANs), to predict internal anatomy from surface data, allowing for the generation of new synthetic images by adjusting internal anatomy parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional data augmentation methods (image rotation, translation, warping) are used to increase training data quantity, then the quantity of training images is improved, but the data augmentation is performed at the image level rather than patient level, leading to insufficient variance and inability to capture different anatomical configurations
Solution Approach 1:
The patent creates synthetic copies of patient anatomy by generating virtual X-ray images from real patient surface data using GANs. Instead of merely transforming existing images, the system synthesizes entirely new anatomical configurations that maintain statistical consistency with real medical data, thereby increasing both quantity and anatomical diversity of training data
Solution Approach 2:
The patent modifies anatomical parameters by adjusting the generated topogram images through controlled transformations that preserve anatomical plausibility. By changing parameters such as organ positions, sizes, and shapes within realistic bounds, the system generates diverse training samples that reflect natural anatomical variations without requiring actual patient exposures
2Quantity of substance
If more X-ray images are acquired for deep learning training, then training data quantity is improved, but patient exposure to radiation increases
Solution Approach 1:
The patent creates synthetic copies of patient anatomy by generating virtual X-ray images from real patient surface data using GANs. Instead of merely transforming existing images, the system synthesizes entirely new anatomical configurations that maintain statistical consistency with real medical data, thereby increasing both quantity and anatomical diversity of training data
Solution Approach 2:
The patent introduces surface data capture and GAN-based generation as an intermediary process between actual X-ray imaging and training data acquisition. This intermediary approach allows the system to obtain anatomical information without direct X-ray exposure, using non-ionizing surface scanning followed by computational synthesis to produce training images
3Adaptability or versatility
If existing data augmentation methods alter images through rotation, translation, and warping, then image variety is improved, but these methods require multiple points of view for a given image and patient, failing to augment data at the patient level
Solution Approach 1:
The patent creates synthetic copies of patient anatomy by generating virtual X-ray images from real patient surface data using GANs. Instead of merely transforming existing images, the system synthesizes entirely new anatomical configurations that maintain statistical consistency with real medical data, thereby increasing both quantity and anatomical diversity of training data
Solution Approach 2:
The patent transitions from 2D image-level transformations to 3D patient-level anatomy generation. By working with volumetric surface data and generating topogram images that represent internal anatomy, the system creates variations in the anatomical dimension rather than merely transforming image perspectives, enabling true patient-level data augmentation
Data Source
AI summary
For anomaly detection based on topogram predication from surface data, a sensor captures the outside surface of a patient. A generative adversarial network (GAN) generates a topogram representing an interior anatomy based on the outside surface of the patient. An X-ray image of the patient is acquired and compared to the generated topogram. By quantifying the difference between the real X-ray image and the predicted one, anatomical anomalies may be detected.