Questions about how individual parts of claim 1 might be changed. A question may describe something different from what the claim language recites, or a version of it; no view is offered on which.
1.2Direct adversarial image synthesis
What ifWhat if a conditional adversarial generator converted the dorsal photograph into a radiograph in one forward pass?
Claim language“the stable diffusion model”
A convolutional generator maps the photograph directly to a grayscale image. During training, a discriminator scores paired photographs and radiographs, while a pixel reconstruction loss ties the generated image to the paired radiograph.
1.4Correlated gaussian latent noise
What ifWhat if the terminal latent noise had a Gaussian distribution with nonzero covariance between neighboring positions?
A fixed spatial filter correlates independent Gaussian samples before the forward process mixes them with the encoded radiograph. The denoiser learns to predict the correlated perturbations, and reverse sampling starts from noise produced by the same filter.
1.5Convolutional photo feature extraction
What ifWhat if a stack of spatial convolutions extracted the dorsal photograph features?
Claim language“a fully connected layer”
Strided 3-by-3 convolutions reduce the photograph to a grid of local feature vectors. The conditioning module reads each grid cell as a separate vector, preserving its row and column position.
1.6Channelwise feature modulation
What ifWhat if channelwise affine modulation injected the dorsal photograph features into the denoiser?
Claim language“via cross attention”
A projection converts the photograph feature vector into one gain and one offset per channel of an intermediate denoiser block. The block multiplies each normalized feature channel by its gain and adds its offset at every spatial position.
1.7Direct pixel space denoising
What ifWhat if the reverse diffusion process generated radiograph pixels directly from a full-resolution noise image?
Claim language“inputting a denoised latent vector into the decoder”
During training, Gaussian noise is added directly to normalized radiograph pixels. At inference, a photo-conditioned denoiser iteratively updates a single-channel noise image, and the final pixel array supplies the image for spinal curvature assessment.
Section 2 · machine-generated
Variations
Further arrangements of the same subject matter, described here as of the publication date at the foot of this page.
Other field of use
Spinal curvature mapping in veterinary imaging
A veterinary imaging workstation pairs dorsal photographs of dogs with dorsoventral spinal radiographs acquired in matched prone postures. During training, corresponding shoulder and pelvic landmarks define a two-dimensional alignment transform, and the encoder maps each aligned radiograph into latent space for noise prediction. At inference, the denoiser uses the dorsal photograph to condition reconstruction, and software measures the angle between selected vertebral endplates in the synthesized image.
Component substitution
A diffusion transformer replaces the convolutional denoiser
A diffusion transformer replaces the U-shaped convolutional denoiser while retaining the radiograph encoder and decoder. The transformer divides each noisy latent into patches, adds spatial position and diffusion timestep embeddings, and uses cross-attention to incorporate dorsal photograph features. A linear output layer predicts latent noise patch by patch, and the sampler reassembles the patches into a noise estimate for each denoising step.
Spatial photograph features replace the global vector
A convolutional image encoder replaces the fully connected layer that compresses the dorsal photograph into one feature vector. The encoder produces a two-dimensional feature grid, and the system flattens the grid into tokens with row and column position embeddings. At intermediate denoiser layers, latent features serve as cross-attention queries and the photograph tokens supply keys and values.
Parameter or scale variation
High resolution synthesis with a fixed sampling schedule
The system fits each aligned image pair onto a canvas 768 pixels wide and 1,536 pixels high using uniform scaling and padding. An encoder with eightfold spatial downsampling produces a latent grid 96 cells wide and 192 cells high, and training samples noise levels from a schedule of 1,000 timesteps. At inference, the sampler applies 40 denoising updates before the decoder reconstructs the image, and angle measurement uses endplate coordinates expressed in the uniformly scaled image space.
Combination with existing systems
Integration with a radiology image archive
An image synthesis service connects to a radiology image archive and retrieves an acquired spinal radiograph through its examination identifier. The service reads the paired dorsal photograph, generates a synthetic image, and stores it as a derived image series under the same examination with a separate series identifier and a provenance record. A review workstation displays the acquired and synthetic series side by side and records the calculated endplate angle for each series.
A landmark detector trained on annotated spinal images produces confidence maps for vertebral endplate corners and assigns each detected vertebra a level. Software fits a line through the corner pair for each endplate, orders the vertebrae by level, and selects upper and lower end vertebrae by their opposing tilts within each detected curve. The measurement module applies the same procedure to the generated and acquired images and records the angles, selected vertebral levels, and absolute angle difference.
Specific version of a part
Landmark driven training registration
The training pipeline marks the cervicothoracic junction and both iliac crests on paired images, using corresponding surface landmarks in the photograph. A similarity transform aligns the pelvic midpoint and shoulder-to-pelvis axis defined by the upper landmark and pelvic midpoint, and a regularized thin-plate spline reduces residual landmark displacement.
Cosine scheduled latent noise
A convolutional encoder maps each aligned grayscale radiograph to a four-channel latent grid at one-eighth of the input width and height. A 1,000-step cosine schedule decreases the retained signal while increasing the contribution of independent unit Gaussian noise across latent coordinates.
Endplate landmark angle comparison
A landmark detector locates the left and right endpoints of each vertebral endplate, and a curve selection routine identifies the upper and lower end vertebrae. The system calculates the angle between the selected endplate lines on each paired image and reports the mean absolute difference in degrees.
Section 3 · quoted from USPTO records
Other work
Passages reproduced word for word from two earlier-published U.S. patent documents, placed beside the language of claim 1. No characterization is offered. Each passage is marked A or B for the document it comes from; ¶ numbers are the paragraph numbers in the USPTO’s published full text of that document.
1.1step 1: inputting the dorsal photograph provided by a user, and ensuring clarity and integrity of a dorsum;
1.2step 2: inputting the dorsal photograph as a prompt into the stable diffusion model, wherein the stable diffusion model includes an encoder, a decoder, and a U-Net, and is capable of converting the dorsal photograph into the spinal radiograph;
1.3step 3: during a model training process, performing alignment processing on the dorsal photograph and the spinal radiograph such that a posture and a shape of the dorsal photograph are consistent with a posture and a shape of the spinal radiograph;
Further, the imaging system may include a synchronization device for correlating a radiograph of said radiographic apparatus with a CBCT image generated by said 4D-CBCT apparatus, such that a reference radiograph can be determined.
B · US 2016/0174921 A1 · abstract
1.4step 4: inputting an aligned spinal radiograph into the encoder to obtain a latent vector, gradually adding noise to the latent vector, and mapping a noisy latent vector to an isotropic Gaussian distribution;
1.5step 5: extracting features of the dorsal photograph using a fully connected layer, and mapping the dorsal photograph into a vector space;
1.6step 6: denoising pure noise using the U-Net, and simultaneously mapping a feature vector of the dorsal photograph to an intermediate layer of the U-Net via cross attention;
1.7step 7: inputting a denoised latent vector into the decoder to generate a spinal radiograph, and predicting a condition of scoliosis; and
1.8step 8: automatically determining a Cobb angle using the generated spinal radiograph, and comparing the Cobb angle with a Cobb angle of an actual spinal radiograph to evaluate effectiveness and accuracy of the method.
For example, the user may evaluate a similarity between the radiographic image and the preview image with respect to the radiographic image obtained by the image obtaining unit 160.
A · US 2014/0037178 A1 · ¶ 85
E. comparing the radiograph generated in step D with the reference radiograph of step A.; F. if the radiograph of step D is sufficiently close to the reference radiograph of step A, determining a position of a tissue of said patient during step D from the CBCT image correlated with said reference radiograph.
B · US 2016/0174921 A1 · ¶ 35
An empty cell means only that no passage was reproduced there.
Statements in these documents
Accordingly, there is a need to reduce exposure of radiation by the radiographic image photographing apparatus.
A · US 2014/0037178 A1 · ¶ 8
The present invention provides a radiographic image photographing apparatus and method capable of simultaneously displaying a parameter value of the radiographic image photographing apparatus and a preview image corresponding to the parameter value, thereby efficiently reducing an amount of radiation to which a human body is exposed.
A · US 2014/0037178 A1 · ¶ 10
Methods for predicting motion of patient's tissues during treatment, such as for instance motion of a tumor or of an organ at risk (OAR), are desired in radiation therapy.
B · US 2016/0174921 A1 · ¶ 3
Predicting motion or position of tissues of patients during treatment is notably necessary in order to avoid radiating OARs and/or in order to irradiate a tumour with a required dose.
B · US 2016/0174921 A1 · ¶ 3
Cite aspatent.fyi, US 2026/0354748 A1, claim 1, first published 2026-11-19T03:31:31-05:00, https://patent.fyi/US20260354748/, SHA-256 5951e0c1250b99c22dcc424d8c9e8e3745101af714f296c78c2eb83e18a03ac6