Eye on AI: How Artificial Intelligence Could Transform Care for Children With Sarcoma 

Sarcomas, a diverse group of cancers that can arise in the bones and soft tissues, account for only about 1% of childhood cancers but encompass more than 100 subtypes. While rare, the stakes are life-altering for each family affected, and pinning down the exact type of sarcoma a child has demands a wide array of tests and analyses as well as careful consideration by specialists. Even then, specialists often disagree, and families without access to major cancer centers can end up waiting weeks while tissue samples are shipped, processed, and reviewed. The whole time, the disease marches on. 

These factors, captured in detail in this earlier blog post, are what make a new artificial intelligence (AI)-based approach to pediatric sarcoma identification so exciting. 

Adam Thiesen, an MD/PhD candidate from UConn Health and The Jackson Laboratory, and colleagues presented a study at the AACR Annual Meeting 2025 in which they tested whether AI can potentially help overcome some of the challenges and provide physicians with a more efficient and reliable way to distinguish between the different forms of the disease, so that they can pair their patients with the right therapies. 

“We aimed to generate tools that improve sarcoma diagnostics and do so in a way that can be implemented at cost and [from] a distance, which could help mitigate against disparities in access to precision oncology care on the basis of treatment setting,” said Thiesen. 

Building a Virtual Pediatric Sarcoma Expert 

Thiesen and his colleagues asked a simple question: Could a computer learn to spot pediatric sarcoma subtypes from the standard hematoxylin-and-eosin (H&E) slides every pathology lab already produces?  

To find out, the team—led by Jill Rubinstein, MD, PhD, of The Jackson Laboratory—assembled 691 digitized images of pathology slides from four children’s hospitals. According to Thiesen, this is likely the largest set of pediatric sarcoma images compiled from U.S. institutions to date.  

The images from the different institutions were then harmonized using open-source software created by Sergii Domanskyi, PhD, and then sliced into tiles. Deep-learning models then extracted numerical “fingerprints” from those tiles, which were then stitched into a single prediction for the whole slide.  

“By digitizing tissue pathology slides, we translated the visual data a pathologist normally studies into numerical data that a computer can analyze,” said Thiesen. “Much like our cell phones can recognize a person’s face in photos and automatically generate an album of photos of that person, our AI-based models recognize certain tumor morphology patterns in the digitized slides and group them into diagnostic categories associated with specific sarcoma subtypes.” 

Two sarcoma categories that frequently affect children and adolescents are bone cancers and soft tissue sarcomas. In particular, osteosarcoma and Ewing sarcoma are the two most common forms of bone cancer diagnosed in young people, whereas the most prevalent soft tissue sarcoma diagnosis is rhabdomyosarcoma. 

Impressively, the AI-driven models accurately identified the sarcoma type in at least 92% of cases when it came to separating Ewing sarcoma from other sarcomas, distinguishing rhabdomyosarcoma from non-rhabdomyosarcoma soft-tissue sarcomas, and even telling the difference between the notoriously similar alveolar rhabdomyosarcoma and embryonal rhabdomyosarcoma.  

These models rival expert review even though they can run on everyday hardware, a deliberate choice to keep the pipeline “lightweight” and accessible. “After the standard data preprocessing, clinicians could theoretically use our models on their own laptops,” Thiesen said. 

Faster Answers, Finer Care 

Diagnostic accuracy has always been a challenge with practical implications, as this case illustrates, and improving diagnostic speed and consistency are where the model could help transform lives. With AI-driven models, the hope is that pediatric sarcoma subtypes, including rare ones, could be correctly diagnosed much sooner, so that future patients can be guided straight to the therapy or clinical trial that might help them most.  

The importance of accurate diagnosis for sarcoma cannot be understated, with one retrospective study published in Clinical Cancer Research finding that being matched to the right therapy significantly increased patient survival. 

“Even with access to the right tools, sarcoma diagnoses have high rates of discordance due to the large variety of subtypes,” Thiesen explained. “Our models could help level the playing field, providing patients with access to quick, streamlined, and highly accurate diagnoses regardless of geographic location or treatment setting.” 

In practice, Thiesen envisions a streamlined workflow where a pathologist can scan the routine H&E slide, upload the digital image to the program, and receive the software’s sarcoma subtype prediction in a timely fashion. His team is now refining the model so that future versions will also flag the exact tissue regions that most influenced the prediction. That added visual feedback would improve transparency and avoid the pitfalls of “black box” AI, helping pathologists decide where to focus any follow-up tests. 

An AI Assistant to Cover Every Child with Sarcoma 

Because pediatric sarcomas are rare, more data will be needed to sharpen the AI model’s discerning eye. The team is inviting other groups to contribute anonymized slides, which could help retrain and refine the model, with the goal of adding support for other rare sarcoma subtypes. “We hope our method inspires others to use similar pipelines,” Thiesen said. 

As more centers contribute image samples and help expand the dataset, the hope is that the models can be refined to deliver precise and timely diagnoses to kids no matter where they live or what form of the disease they have. Future versions of the model could potentially pair the image data with genomic information whenever morphology alone is ambiguous.