AI Tool Reduces the Need for Large Datasets in Medical Image Segmentation

A newly developed artificial intelligence (AI) tool aims to simplify and reduce the costs associated with training medical imaging software, particularly when only a small number of patient scans are available. This AI tool enhances the process of medical image segmentation, in which every pixel of an image is labeled to identify its characteristics, such as distinguishing between cancerous and healthy tissue.

Medical image segmentation is typically performed by highly trained professionals, with deep learning offering a promising method to automate this labor-intensive task. However, deep learning techniques often require vast amounts of annotated image data to function effectively, which can be both expensive and time-consuming to produce. In many clinical settings, such datasets are either unavailable or prohibitively difficult to create.

To address this challenge, a research team led by Li Zhang, a PhD student at the University of California San Diego’s Department of Electrical and Computer Engineering, developed an AI tool capable of learning from only a small number of annotated images. The tool can reduce the data requirements by up to 20 times, potentially making medical imaging more accessible and affordable, particularly in resource-limited environments.

The team's research, published in Nature Communications, demonstrates the effectiveness of the tool across a variety of medical imaging tasks. These tasks include identifying skin lesions in dermoscopy images, detecting breast cancer in ultrasound scans, mapping placental vessels in fetoscopic images, locating polyps in colonoscopy images, and recognizing foot ulcers in standard camera photos. The tool was also extended to work with 3D medical images, such as those used in brain and liver mapping.

In scenarios where annotated data were scarce, the AI tool improved model performance by 10 to 20% compared to existing methods. It also required significantly less training data—eight to 20 times fewer real-world examples—while maintaining or surpassing the performance of traditional methods.

One potential application of this technology is in dermatology, where the tool could assist in diagnosing skin cancer. For example, a trained expert might annotate just 40 images, which the AI system could then use to identify suspicious lesions in real-time during dermoscopy exams. This could potentially lead to faster and more accurate diagnoses, ultimately improving patient outcomes.

The AI tool operates in several stages. First, it learns to generate synthetic images based on segmentation masks, which are color-coded overlays that indicate the areas of an image that correspond to specific features, such as diseased or healthy tissue. Next, the system uses these synthetic images, along with a small set of real annotated examples, to train a segmentation model. A continuous feedback loop ensures that the synthetic data are tailored to enhance the model’s segmentation abilities.

According to Zhang, integrating data generation and model training into a single process is key to the tool’s success. The segmentation performance guides the generation of synthetic data, ensuring that the produced data are not only realistic but also optimized to improve the model’s ability to perform image segmentation tasks.

The research team plans to continue enhancing the AI tool, making it more intelligent and adaptable to different medical contexts. Future developments may also include direct feedback from clinicians, ensuring that the generated data meet the specific needs of real-world medical applications.

Reference: Zhang L, Jindal B, Alaa A, et al. Generative AI enables medical image segmentation in ultra low-data regimes. Nat Commun. 2025. doi:10.1038/s41467-025-61754-6

This article has been republished from the following materials. Note: material may have been edited for length and content. For further information, please contact the cited source. Our press release publishing policy can be accessed here.

This content includes text that has been generated with the assistance of AI. Technology Networks' AI policy can be found here.