It sounds like a pastry but it’s really part of a new wave in scientific imaging. Empanada is the name of an artificial intelligence (AI) program trained on a vast set of data for more rapid and accurate 3-D rendering of subcellular structures. Interested investigators can download it for free

AI program-generated image from kidney
Mitochondrial segmentation of kidney distal tubule using Mitonet, generalist deep learning model that automatically segments mitochondrial instances using Empanada and a highly heterogeneous dataset of labeled mitochondria.

“At the snap of a finger, you can identify all of the mitochondria in a cell,” said Kedar Narayan, Ph.D., senior scientist and group leader of the National Cancer Institute’s Center for Molecular Microscopy at the Frederick National Laboratory (FNL).

The initial release of Empanada focuses on mitochondria, the powerhouses of living cells that produce energy. Subsequent iterations will be tailored to identify other cellular components, such as nuclei, endosomes, and lamellar bodies. 

The textbook diagram of mitochondria shows a plump sausage shape that has been sliced lengthwise in a cross-section that reveals in detail its interior structures. In actuality, however, mitochondria come in different shapes and sizes and can be densely packed, a potential hodge-podge of organelles that can be difficult to recognize individually.  

Moreover, a two-dimensional cross-section of a cell might reveal only some of the hundreds or thousands of mitochondria that are present, depending on where the cross-section was cut. Three-dimensional volume electron microscopy can capture all organelles in a cell, but it results in gigabytes of data, which takes time to process and sort for individual mitochondria. Empanada does this rapidly and with precision. 

“The backbone of Empanada is the data on which it is trained,” Narayan said. “Ours is curated from datasets equivalent to 2 petabytes of data. Our model sampled every publicly available volume electron microscopy dataset, and we distilled it down to about 1.5 million images. That means you are training AI with vastly heterogenous data. That’s why the AI is so powerful. No model is perfect, but it does perform exceedingly well on any number of different cell types.” 

The technology was developed at FNL, a national resource for collaboration, innovation, and technology dissemination. The AI software comes as a plug-in with a user-friendly interface and does not require data analysis expertise to get results.  

The current release of Empanada is available for download at no cost.  Authored by FNL research associates Abhishek Bhardwaj and Madeline Berry from Narayan’s group, the software package will perform rapid segmentation on both two- and three-dimensional electron microscopy images. 

Media Inquiries

Mary Ellen Hackett
Manager, Communications Office