An international collaboration with FGV EMAp used computer vision and artificial intelligence techniques to predict the tenderness and fat content of beef and pork based

Study paves the way for the development of apps that help consumers identify, through images, more tender and juicy cuts of meat | Image: Envato

Brazilians love gathering family and friends for a good barbecue. But when it comes to choosing the meat, the same question always arises: what's the best cut to guarantee a perfect experience? This decision, usually made by eye, may soon get a little help from science.

Researchers from three countries - Brazil, Canada, and the United States - have developed an artificial intelligence - based system capable of predicting the tenderness and intramuscular fat of raw meat using only photos taken with a smartphone. The study, published in the journal Meat Science, involved Dário Oliveira, a researcher from the School of Applied Mathematics at Fundação Getulio Vargas (FGV EMAp).

“This kind of international collaboration is essential for FGV EMAp and for advancing data science in Brazil. We brought together different areas of expertise to solve a real problem involving technological innovation and direct impact on people’s lives,” says Dário.

How was the research conducted?

The AI model was trained using images of 924 beef steaks and 514 pork cuts, all captured with a smartphone under controlled lighting conditions. The images were collected over a year by meat quality expert Márcio Duarte, then working at the University of Guelph, Canada.

With this dataset, researcher João Dórea from the University of Wisconsin–Madison developed a project led by Guilherme Lobato Menezes, in collaboration with Dário Oliveira, to train neural networks to analyze and predict meat quality from the images. Each photo was matched with lab-based data, such as shear force - which measures how much force is needed to cut the meat - and the percentage of fat between muscle fibers, known as intramuscular fat (IMF).

In addition to classifying cuts into categories like “tender,” “intermediate,” or “tough,” the researchers applied regression tests to predict exact numerical values for tenderness and fat content. This approach allowed the model to estimate, from an image, measurements traditionally obtained only through lab testing.

In practice, 1,000 random pairs were generated from the 142 images in the test set. For each pair, the system estimated shear force and fat percentage, then indicated which cut would theoretically be more tender or have a higher intramuscular fat content. This allowed the researchers to evaluate how consistent the model would be in situations similar to what a consumer might face when choosing between two cuts of meat at the supermarket.

Using this method, the model achieved 76.5% accuracy in predicting beef tenderness - significantly outperforming the 130 consumers who took part in the experiment. Given the same images used by the model, these participants attempted to identify the more tender steak in each pair but got it right only 46.7% of the time on average. In the case of pork, due to the high similarity between cuts, the images were not shown to consumers. Even so, the model selected the more tender cut in 81.5% of cases. As for intramuscular fat, the model reached 77% accuracy for beef and 79% for pork.

For researcher Guilherme, being able to estimate intramuscular fat with precision is a major step forward.

“It directly influences the meat’s juiciness and, consequently, its flavor. Our research allows us to quantitatively predict that percentage, enabling consumers to be much more specific about their preferred fat level, something impossible to assess accurately just by looking,” he explains.

Researchers behind the study: João Dórea, Márcio Duarte, Dário Oliveira, and Guilherme Menezes (top to bottom, left to right) | Image: Personal archive

Science applied to the meat supply chain

For now, the system works only for two types of cuts: beef striploin and pork loin. The researchers plan to test the model in the future with other cuts, under different lighting conditions, and using meat from various breeds and origins. In addition, they hope the study will serve as a foundation for a startup or tech company to develop an app that helps consumers compare meat cuts at the time of purchase. According to the researchers, the same tool could also be valuable for the industry, enabling more fair and transparent pricing based on the actual quality of the product.

“For a country like Brazil, which is among the largest meat producers and exporters in the world, agribusiness is a strategic sector. And a tool like this can transform the entire chain of production, trade, and consumption,” emphasizes Dário.