Noise preserving models and methods for resolution recovery of x-ray computed tomography images
Inventors
Melnyk, Roman • NAGARE, Madhuri Mahendra • Tang, Jie • Rahman, Obaidullah • Nett, Brian E • Sauer, Ken • Bouman, JR., Charles Addison
Assignees
University of Notre Dame • Purdue Research Foundation • GE Precision Healthcare LLC
Publication Number
US-12299849-B2
Publication Date
2025-05-13
Expiration Date
2042-06-20
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Noise preserving models and methods for resolution recovery of x-ray computed tomography (e.g., using a computerized tool) are enabled. For example, a system can comprise: a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory, wherein the computer executable components comprise: a pair generation component that generates a pair of images, the pair of images comprising an input image and a ground truth image, a training component that trains a machine learning based sharpening algorithm by approximately minimizing a loss function that determines an error between a sharpened image and the ground truth image, and a sharpening component that, using the sharpening algorithm, sharpens the input image to generate the sharpened image, wherein the sharpened image comprises a second noise that is similar in intensity to a first noise of the input image.
Core Innovation
The invention relates to noise preserving models and methods for resolution recovery of x-ray computed tomography (CT) images using machine learning techniques. The described system generates image pairs consisting of an input image (which may be blurred and noisy) and a ground truth image (which is noise-free or less noisy). During training, a controlled amount of noise, sampled from physical sources such as water phantoms, is added independently and in a scaled manner to both the input and ground truth images, ensuring the resulting sharpened images preserve a noise intensity similar to the input image.
A key aspect of the method involves training a machine learning-based sharpening algorithm that minimizes a loss function measuring the difference between the sharpened image and the ground truth. Adjustable parameters allow the user to fine-tune the degree of sharpening, the amount of noise added, and the balance between preserving noise texture and enhancing resolution. The solution further includes processes such as edge magnitude-based blending and corrections based on Hounsfield Unit values to reduce artifacts in the resulting images, ensuring clinical relevance by retaining essential image details while preventing overcorrection, especially at soft tissue boundaries.
The problem addressed by this invention is that existing methods for sharpening CT images often amplify noise or result in overly smoothed outputs that suppress clinically important texture and detail. Previous approaches fail to preserve the input noise characteristics, which are preferred by radiologists, or introduce artifacts and excessive noise when sharpening high-resolution images. This invention overcomes these deficiencies by providing a noise-preserving sharpening filter (NPSF) that yields images with enhanced resolution and preserved noise energy and texture, using adjustable scaling and training configurations.
Claims Coverage
The patent claims three primary inventive features across its independent claims, covering system, method, and non-transitory machine-readable storage medium aspects.
Noise-preserving training data generation for sharpening algorithms
The system generates a first set of pairs of training images from a second set where each pair includes an input image and a noise-free ground truth image, the input image being a blurred version of the ground truth. For each pair, noise sampled from an actual source is added to both images, with the added noise in each scaled separately so that the noise in the resulting training input and ground truth images is substantially similar in intensity (according to a defined threshold). This maintains a consistent noise level for the sharpening task.
Machine learning-based sharpening with noise preservation
A machine learning-based sharpening algorithm is trained using these noise-preserved training pairs by minimizing a loss function that measures image blur reduction while maintaining a substantially similar noise intensity between the training input and ground truth images. When applied to new input images, the sharpening algorithm produces sharpened outputs with reduced blur and a noise intensity substantially similar to the original input, according to the defined threshold.
Integration into system, method, and machine-readable storage medium
The inventive features, including noise-scaled training data generation and noise-preserving sharpening, are implemented as: (1) a system comprising memory and processor configured to perform these operations, (2) a computer-implemented method with similar steps for generating training data, training, and sharpening, and (3) a non-transitory machine-readable storage medium containing executable instructions for deploying the approach on suitable computing platforms.
The independent claims secure protection for the generation of noise-matched training pairs, training and deployment of a sharpening algorithm that preserves noise characteristics, and coverage for these techniques as a system, method, and software implementation.
Stated Advantages
Enables sharpening of medical CT images without substantially increasing or reducing the intensity of noise, thereby preserving noise energy and texture preferred for clinical interpretation.
Reduces or avoids artifacts typically associated with conventional sharpening, including overcorrection in soft tissue areas, through processes like edge magnitude-based blending.
Provides adjustable parameters for balancing resolution improvement and noise, offering flexibility to achieve clinically desirable image quality.
Permits improved resolution of high-density bone structures while maintaining image noise at levels consistent with clinician preferences.
Documented Applications
Noise-preserving sharpening and resolution recovery of x-ray computed tomography (CT) images for medical diagnostics.
Sharpening and noise control in images from other medical modalities, including radiation therapy images, X-ray images, digital radiography, X-ray angiography, panoramic X-ray, mammography (including tomosynthesis), magnetic resonance imaging (MRI), ultrasound, color flow doppler, positron emission tomography (PET), single-photon emissions computed tomography (SPECT), nuclear medicine, synthetic and augmented medical images, and three-dimensional medical imaging.
Interested in licensing this patent?