EOS Imaging has filed a patent for a medical imaging conversion method that automatically converts real x-ray images of a patient into digitally reconstructed radiographs (DRR). The method uses a convolutional neural network (CNN) or a group of CNNs to differentiate between anatomical structures and generate DRRs representing specific structures. GlobalData’s report on EOS Imaging gives a 360-degree view of the company including its patenting strategy. Buy the report here.
According to GlobalData’s company profile on EOS Imaging, single photon emission computed tomography (SPECT) was a key innovation area identified from patents. EOS Imaging's grant share as of June 2023 was 1%. Grant share is based on the ratio of number of grants to total number of patents.
Medical imaging conversion method using convolutional neural networks
A recently filed patent (Publication Number: US20230177748A1) describes a medical imaging conversion method that automatically converts real x-ray images of a patient into digitally reconstructed radiographs (DRR) using convolutional neural networks (CNNs). The method aims to differentiate between different anatomical structures and generate DRRs that represent specific structures without including others.
The first claim of the patent describes the conversion of real x-ray images into DRRs representing the first anatomical structure without representing the second anatomical structure. This conversion is achieved through a single operation using either one CNN or a group of CNNs that are trained to differentiate between the two structures and convert the x-ray images into DRRs.
The second claim extends the method to also convert the real x-ray image into DRRs representing the second anatomical structure without representing the first anatomical structure. This is done using the same single operation and trained CNNs.
The third claim introduces the conversion of real x-ray images into both a first and a second DRR representing the first and second anatomical structures, respectively. Again, this is achieved through a single operation using either one CNN or a group of CNNs that are trained to differentiate between the two structures and convert the x-ray images into DRRs.
The patent also mentions the use of a single generative adversarial network (GAN) as the CNN or group of CNNs in claim 4. Specifically, the GAN can be a U-Net GAN or a Residual-Net GAN.
The method described in the patent is applicable to various anatomical structures, including contiguous vertebrae, regions of the patient's spine, patient hip, lower limbs, knee, shoulder, and rib cage.
The patent further discusses the use of training groups and subsets of x-ray images and corresponding DRRs to train the CNN or group of CNNs. It also mentions the possibility of using frontal and lateral real x-ray images to generate corresponding DRRs.
Additionally, the patent describes the personalization of a 3D model using the medical imaging conversion method. A generic 3D model is used to generate DRRs from frontal and lateral views of the patient, which are then mapped with the DRRs obtained from the real x-ray images to create a patient-specific 3D model.
Overall, the patent presents a method for automatically converting real x-ray images into DRRs using CNNs, allowing for the representation of specific anatomical structures without including others. The method has potential applications in medical imaging and 3D model personalization.