Biomedical imaging, including methods such as magnetic resonance imaging (MRI), computed tomography, and ultrasound, plays a significant role in medical diagnostics and research. Semantic segmentation, an important technique within this field, involves partitioning images into meaningful segments and assigning labels to each pixel based on the object or region it represents. This segmentation provides detailed maps of various structures, such as different tissue types and anomalies like cancerous tissues.
Accurate semantic segmentation relies on understanding the contextual and spatial relations within images. When automated it often relies on convolutional neural networks (CNNs), advanced machine learning algorithms adapted to recognise patterns in images. One of the most notable CNN architectures currently used for biomedical imaging tasks is U-Net, named due to its U-shaped architecture.
CNNs are well-suited for semantic segmentation due to their ability to automatically learn hierarchical features from raw pixel data. Through the usage of multiple layers of convolution and pooling operations, CNNs extract abstract representations of images, capturing both low-level details and high-level contextual information. This enables CNNs to accurately localise objects and define their boundaries. Leveraging CNNs for semantic segmentation however typically requires manual configuration of hyperparameters, which modify the model’s architecture and learning process. This manual tuning poses a significant barrier, particularly for non-experts, limiting the technology’s broader usability.
To address this challenge, researchers at Heidelberg, Germany, developed nnUNet in 2020, an automated method for configuring U-Net for image segmentation tasks. nnU-Net streamlines the segmentation process by automatically adjusting parameters, and settings based on the provided dataset, eliminating the need for manual tuning. It also modifies the dataset itself to extract all the informational value from it by normalising, reshaping and if needed upsizing it. This self-configuring approach enhances adaptability to diverse datasets and imaging modalities, making segmentation more accessible. Demonstrating its efficacy across various biomedical imaging tasks, including MRI analysis and cell microscopy, nnU-Net has surpassed many existing approaches, highlighting its potential to enhance image analysis in healthcare and research. Its success, underscored by the widespread adoption and subsequent citations, reflects a broader trend in the field towards democratising image analysis and making advanced techniques accessible to all.
Since U-Net’s inception in 2015, there have been over 18000 citations to that first paper. Based on U-Net most researchers have opted to design their own architecture specific to their needs. The proliferation of research and technologies like nnU-Net however suggests a growing audience of more casual people interested in the field, a clear milestone in this maturing field.
At CONcISE, our research tackles increasingly complex imaging problems, and while its results will not be as easily applicable to a wider audience, there is a hope it will inspire further developments in creating user-friendly solutions for increasingly intricate challenges. As these technologies evolve, they hold the promise of unlocking new insights and discoveries in biomedical imaging, thus benefiting society.
