Medical AI often faces a core challenge: high-quality labeled medical data are costly to acquire, and collecting millions of images isn’t always feasible. For image segmentation, where precise marking of specific organs or structures is required, acquiring enough training samples can be a major hurdle. This patent explores whether an AI can better utilize limited examples by learning two related tasks simultaneously. It employs a multi-task adversarial network where the generator learns both organ segmentation and bone-suppression segmentation at the same time. The system features a U-Net-based generator with traditional and dilated convolution layers, refined through adversarial training. The rationale for learning both tasks is that related tasks can offer complementary insights. Instead of just relying on a few labeled images for one goal, joint learning provides the model with another structured problem, helping it extract more useful features. The patent addresses the challenge of achieving accurate medical-image segmentation with limited training data. Its importance extends beyond a specific architecture, illustrating a different approach to data scarcity: making each available example teach the network multiple things. This principle can be especially valuable in fields where expert annotations are costly or limited. In medicine, where expanding datasets isn’t as straightforward as in general image collections, enhancing what AI can learn from each labeled example may be just as crucial as increasing the number of examples.
Learning more from fewer medical images
Patent number: CN 114764805 A
Inventor(s): W. Huo and X. Tian
Citation: W. Huo and X. Tian, “一种图像分割方法、分割装置、终端设备及存储介质 [Image-segmentation method, segmentation apparatus, terminal device and storage medium],” China Patent Application CN 114764805 A, Jul. 19, 2022.
