Canon Medical Systems Corporation

University of California

University Health Network

Chung Chan

Evren Asma

Jian Zhou

Li Yang

Wenyuan Qi

Zhou Yu

Yi Qiang

Bernice Hoppel

Jeffrey Kolthammer

Patrik Rogalla

Kurt Walter Schultz

Tiantian Li

Zhaoheng Xie

A method, system, and computer readable medium to perform nuclear medicine scatter correction estimation, sinogram estimation and image reconstruction from emission and attenuation correction data using deep convolutional neural networks. In one embodiment, a Deep Convolutional Neural network (DCNN) is used, although multiple neural networks can be used (e.g., for angle-specific processing). In one embodiment, a scatter sinogram is directly estimated using a DCNN from emission and attenuation correction data. In another embodiment a DCNN is used to estimate a scatter-corrected image and then the scatter sinogram is computed by a forward projection.

Scatter estimation for PET from image-based convolutional neural network

A deep learning (DL) convolution neural network (CNN) reduces noise in positron emission tomography (PET) images, and is trained using a range of noise levels for the low-quality images having high noise in the training dataset to produce uniform high-quality images having low noise, independently of the noise level of the input image. The DL-CNN network can be implemented by slicing a three-dimensional (3D) PET image into 2D slices along transaxial, coronal, and sagittal planes, using three separate 2D CNN networks for each respective plane, and averaging the outputs from these three separate 2D CNN networks. Feature-oriented training can be implemented by segmenting each training image into lesion and background regions, and, in the loss function, applying greater weights to voxels in the lesion region. Other medical images (e.g. MRI and CT) can be used to enhance resolution of the PET images and provide partial volume corrections.

Apparatus and method for medical image reconstruction using deep learning to improve image quality in positron emission tomography (PET)

Existing, low quality images can be restored using reconstruction or a combination of post-reconstruction techniques to generate a real patient phantom. The real patient phantom (RPP) can then be simulated in Monte Carlo simulations of a higher performance system and a lower performance system. Alternatively, the RPP can be simulated in the higher performance system, and a real scan can be performed by an existing, lower performance system. The higher performance system can be differentiated from the lower performance system in a variety of ways, including a higher resolution time of flight measurement capability, a greater sensitivity, smaller detector crystals, or less scattering. A neural network can be trained using the images produce by the higher performance system as the target, and the images produced by the lower performance system as the input. After training, the trained neural network can be used to output input images taken in a lower performance system with higher performance system characteristics.

Neural network for improved performance of medical imaging systems

A system and method for training a neural network to denoise images. One noise realization is paired to an ensemble of training-ready noise realizations, and fed into a neural network for training. Training datasets can also be retrospectively generated based on existing patient studies to increase the number of training datasets.

Noise to noise ensemble learning for pet and CT image and data denoising

A deep learning (DL) convolution neural network (CNN) reduces noise in positron emission tomography (PET) images, and is trained using a range of noise levels for the low-quality images having high noise in the training dataset to produceuniform high-quality images having low noise, independently of the noise level of the input image. The DL-CNN network can be implemented by slicing a three-dimensional (3D) PET image into 2D slices along transaxial, coronal, and sagittal planes, using three separate 2D CNN networks for each respective plane, and averaging the outputs from these three separate 2D CNN networks. Feature-oriented training can be implemented by segmenting each training image into lesion and background regions, and, in the loss function, applying greater weights to voxels in the lesion region. Other medical images (e.g. MRI and CT) can be used to enhance resolution of the PET images and provide partial volume corrections.

Apparatus and method for medical image reconstruction using deep learning to improve image quality in position emission tomography (PET)

A method and apparatus is provided to iteratively reconstruct an image from gamma-ray emission data by optimizing an objective function with a spatially-varying regularization term. The image is reconstructed using a regularization term that varies spatially based on an activity-level map to spatially vary the regularization term in the objective function. For example, more smoothing (or less edge-preserving) can be imposed where the activity is lower. The activity-level map can be used to calculate a spatially-varying smoothing parameter and/or spatially-varying edge-preserving parameter. The smoothing parameter can be a regularization parameter &#x3b2; that scales/weights the regularization term relative to a data fidelity term of the objective function, and the regularization parameter &#x3b2; can depend on a sensitivity parameter. The edge-preserving parameter &#x3b2; can control the shape of a potential function that is applied as a penalty in the regularization term of the objective function.

Activity-dependent, spatially-varying regularization parameter design for regularized image reconstruction

A method and apparatus is provided to correct for scatter in a positron emission tomography (PET) scanner, the scatter coming from both within and without a field of view (FOV) for true coincidences. For a region of interest (ROI), the outside-the-FOV scatter correction are based on attenuation maps and activity distributions estimated from short PET scans of extended regions adjacent to the ROI. Further, in a PET/CT scanner, these short PET scans can be accompanied by low-dose X-ray computed tomography (CT) scans in the extended regions. The use of short PET scans, rather than full PET scans, provides sufficient accuracy for outside-the-FOV scatter corrections with the advantages of a lower radiation dose (e.g., low-dose CT) and requiring less time. In the absence of low-dose CT scans, an atlas of attenuation maps or a joint-estimation method can be used to estimate the attenuation maps for the extended regions.

Method and apparatus for scatter correction in position emission tomography (PET) imaging by performing a short PET scan in an extended region to estimate scatter coming from outside of the field of view (FOV)

A method and apparatus is provided to iteratively reconstruct a computed tomography (CT) image using a spatially-varying content-oriented regularization parameter, thereby achieving uniform statistical properties within respective organs/regions and different statistical properties (e.g., degree of smoothing and noise level) among the respective organs/regions. For example, less smoothing and sharper features/resolution can be applied within a lung region than within a soft-tissue region by using a smaller regularization parameter value in the lung region than in the soft-tissue region. This can be achieved, e.g., using a minimum intensity projection to suppress/eliminate sub-solid nodules in the lung region. The content-oriented regularization parameter can be generated by reconstructing an initial CT image, which is then segmented/classified according to organs and/or tissue type. Segmenting the image and generating the content-oriented regularization parameter can be integrated into one process by applying an HU-to-&#x3b2; mapping to the CT image.

Apparatus and method for context-oriented iterative reconstruction for computed tomography (CT)

A method and apparatus is provided generate a display image that optimize a tradeoff between resolution and noise by using blending weights/ratio based on the content/context of the image. The blending weights control the relative weights when combining multiple computed tomography (CT) images having different degrees of smoothing/denoising to generate a display image having the optimal tradeoff lying within the continuum between/among the CT images. The blending weights are automated based on information indicating the content/context of the display image (e.g., the segmented tissue type, average attenuation, and the display setting such as window width and window level). Thus, indicia indicating content/context of the image determine the weighting coefficients, which are used in a weighted sum, e.g., to combine the plurality of images with different noise/smoothing parameters into a single blended image, which is displayed.

Growing community of inventors

Glenview, IL, United States of America

Chung Chan