AI Method for Sharpening Low-Dose PET Scans with Multi-Task Learning

This patent describes an AI method using a special type of neural network called a multi-task learning generative adversarial network (GAN) to improve the quality of medical images from low-dose PET scans, making them look like standard-dose scans.

What it covers

The patent describes a method for creating an AI system that can take blurry, low-dose Positron Emission Tomography (PET) images and make them clearer, similar to standard-dose PET images. It starts by building a "U-Net type picture generator" by connecting an "encoder" and a "decoder" using "skip connections" (Claim 1). Then, it creates multiple "generative adversarial networks" (GANs) by pairing these picture generators with "discriminators" (Claim 1). These GANs learn in parallel, sharing basic information, forming a "first multi-task learning-type generative adversarial network." The system is trained using a "joint loss function" that combines L1 type loss and cross entropy loss, comparing the AI's output to high-quality standard-dose PET images (Claims 1, 2). For example, if a patient receives a low-dose PET scan to reduce radiation exposure, this AI can process those images to produce diagnostic quality results.

What it doesn't cover

• —Does not cover methods for reconstructing PET images without using a "U-Net type picture generator" with "skip connections" (Claim 1).
• —Does not cover systems that use only a single generative adversarial network, as it specifies "a group of generative adversarial networks" and "plurality of picture generators" (Claim 1).
• —Does not cover methods where the "plurality of discriminators" do not use an "input modality," a "tag image," and an "output result" as input (Claim 1).
• —Does not cover methods that do not combine low-dose PET and MR images as input modalities for each GAN (Claim 1).
• —Does not cover training methods that do not use a "joint loss function 1" combining L1 type loss and cross entropy loss functions (Claim 1, 2).

The clever bit

The novelty lies in combining multiple "U-Net type picture generators" with "discriminators" into a "multi-task learning-type generative adversarial network," where these networks learn in parallel and share information. This parallel learning, guided by a specific "joint loss function" that considers both L1 and cross-entropy losses, allows the system to effectively reconstruct high-quality images from very low-dose PET data, leveraging additional MR image input.

Why it matters

Positron Emission Tomography (PET) scans are crucial for diagnosing diseases like cancer, but they expose patients to radiation. This patent matters because it offers a way to get high-quality images from lower radiation doses, making PET scans safer and potentially more accessible. By improving image quality from low-dose data, it could lead to earlier and more accurate diagnoses with reduced patient risk.

Real-world examples

• 1.Medical imaging software for PET scanners
• 2.AI-powered diagnostic tools in oncology
• 3.Radiology workstations
• 4.Research platforms for medical image enhancement