Improving Medical Image Segmentation with AI and Fake Data

This patent describes a method to make AI better at finding specific things in medical images, like tumors, by using computer-generated fake images to train the AI and a special network design.

What it covers

This patent details a method for more accurately segmenting, or outlining, specific areas in medical images using a type of artificial intelligence called a U-Net. First, it takes existing medical images and their manual outlines (called 'real segmentation maps') and feeds them into a Generative Adversarial Network (GAN). This GAN creates new, realistic 'composite images' that look like real medical images but are computer-generated, helping to expand the training data (claim 1, step 3). This expanded dataset is then used to train an 'improved multi-feature fusion segmentation network,' which is a modified U-Net (claim 1, step 5). A key improvement is the addition of a 'Dilated Convolution Module' between the network's shallow and deep feature connections, which helps the network better combine fine details with broader context, making it more adaptable to different target sizes and improving accuracy (claim 1, step 5). Finally, this trained network can accurately segment new medical images, for example, precisely outlining a tumor in an MRI scan.

What it doesn't cover

• —Does not cover medical image segmentation methods that do not use a Generative Adversarial Network (GAN) for data enhancement.
• —Does not cover U-Net based segmentation networks that lack a Dilated Convolution Module in their skip connections.
• —Does not cover image segmentation tasks for non-medical images, such as segmenting objects in photographs or satellite imagery.
• —Does not cover medical image segmentation techniques that do not rely on a U-Net or a multi-feature fusion segmentation network architecture.
• —Does not cover methods that do not use a Dice loss function for training the segmentation network, as specified in claim 4.

The clever bit

The cleverness lies in two main parts: first, using a Generative Adversarial Network (GAN) to create more training data, which helps prevent the AI from 'overfitting' to a small dataset. Second, it integrates a 'Dilated Convolution Module' into the U-Net's skip connections, allowing the network to see objects at different scales more effectively, thus improving its ability to find and outline targets of varying sizes.

Why it matters

Accurate medical image segmentation is crucial for diagnosing diseases, planning treatments, and monitoring patient progress. This method aims to improve the precision of these tasks by addressing common challenges in AI training, such as limited access to diverse medical datasets and the need for robust feature recognition. By generating synthetic data and enhancing the network's ability to fuse different levels of detail, it helps medical professionals get clearer and more reliable outlines of areas of interest, like organs or lesions.

Real-world examples

• 1.Automated tumor detection in MRI or CT scans
• 2.Precise outlining of organs for radiation therapy planning
• 3.Quantifying lesion volumes in neurological scans
• 4.Identifying specific tissue types in pathology slides
• 5.Segmenting blood vessels in angiography images