With more than 50 different subtypes, pediatric soft-tissue sarcomas represent a broad category of disease. And given the rarity of these sarcomas, “it is difficult for pathologists to see a sufficient volume to gain expertise across all variants,” stated Adam Thiesen, BA, MD/PhD Candidate at UConn Health and The Jackson Laboratory, Farmington, Connecticut. “The rise in digitization of histopathology slides has opened the door for applying artificial intelligence [AI] algorithms to this task,” he added. During a presentation at the 2025 American Association for Cancer Research (AACR) Annual Meeting, Mr. Thiesen shared study results on behalf of his colleagues on the use of AI-based models to accurately classify pediatric sarcomas using digital pathology slides alone.1
“Our findings demonstrate that AI-based models can accurately diagnose various subtypes of pediatric sarcoma using only routine pathology images.”— Adam Thiesen, BA
Tweet this quote
“By digitizing tissue pathology slides, we translated the visual data a pathologist normally studies into numerical data that a computer can analyze,” he explained. “Our AI-based models recognize certain tumor morphology patterns in the digitized slides and group them into diagnostic categories associated with specific sarcoma subtypes.”
Study Details
Thus far, more than 800 images have been collected across the Children’s Oncology Group (n = 582), Yale School of Medicine (n = 153), St. Jude’s Children’s Hospital (n = 71), and Massachusetts General Hospital (n = 63). Of the nine sarcoma subtypes represented, a total of 442 images were of embryonal rhabdomyosarcoma and 135 were of alveolar rhabdomyosarcoma. According to Mr. Thiesen, the diversity of the data set was also seen in terms of age, race, and sex of the patients.
“The representation of different subtypes present in our data set allows us to train models on rare subtypes not usually assessed due to lack of data,” he explained. “In addition, the heterogeneity in our data set allows our models to have a holistic view across the spectrum of disease.”
Briefly, the researchers developed and applied open-source software to harmonize the images collected from these different institutions to account for variation in format, staining, and magnification, among other factors. The harmonized images were then converted into small tiles before being fed into deep learning models, which extracted numerical data for analysis by a novel statistical method. Then, summaries of each slide’s features were generated, and trained AI algorithms categorized each slide as a specific subtype. Finally, the investigators developed spatially resolved attention maps, which provide interpretability for the regions of a slide that contain malignant cells.
Key Results
In validation experiments, the AI algorithms identified sarcoma subtypes with high accuracy, according to Mr. Thiesen. Specifically, the AI-driven models correctly distinguished between Ewing sarcoma and other sarcoma types in 92.2% of cases, with an area under the receiver operating characteristic curve (AUROC) of 0.966. The models distinguished correctly between nonrhabdomyosarcoma and rhabdomyosarcoma soft-tissue sarcomas in 93.8% of cases, with an AUROC of 0.969. In terms of rhabdomyosarcomas, the models distinguished between alveolar and embryonal in 95.1% of cases, with an AUROC of 0.95. In addition, despite uneven sample representation, the AI-driven models distinguished among alveolar, embryonal, and spindle cell rhabdomyosarcomas in 87.3% of cases, with an AUROC of 0.88 for alveolar vs embryonal vs spindle cell.
Clinical Implications
“Our findings demonstrate that AI-based models can accurately diagnose various subtypes of pediatric sarcoma using only routine pathology images,” said Mr. Thiesen. “These AI-driven models could help provide more pediatric patients access to quick, streamlined, and highly accurate cancer diagnoses, regardless of their geographic location or health-care setting.”
The classification performance of these AI-based models was reported to be on a par with manual pathology review. According to Mr. Thiesen, unlike with standard clinical histopathology workflow, with these models, no extra stains, immunohistochemistry, or genomics are needed, as they use only hematoxylin and eosin–stained slides. In addition, this approach seems to mitigate disparity in access to precision cancer diagnostics, he added.
There are others benefits with these novel models compared with prior models, Mr. Thiesen noted. First, their models require a fraction of the computing power of previously published models,2 he said. Second, their models are built to allow new images to be added and trained with minimal computational equipment. Third, these novel models can be used on standard clinical laptops, which increases the accessibility to deep learning methods globally.
“To our knowledge, we have the largest general sarcoma data set at a multi-institutional level, with broad representation across subtypes, anatomic locations, race, and sex. We’re very excited about extending this to additional medical centers and capturing even more rare subtypes.”
DISCLOSURE: This research was supported by the National Institutes of Health, The Jackson Laboratory, and Hartford Hospital. The study was organized by surgical oncologist Jill Rubinstein, MD, PhD, Senior Research Scientist at The Jackson Laboratory, and used software created by Sergii Domanskyi, PhD, Associate Computational Scientist at The Jackson Laboratory. Mr. Thiesen reported no conflicts of interest.
REFERENCES
1. Thiesen A, Domanskyi S, Foroughi Pour A, et al: Artificial intelligence for digital pathology and spatial molecular technologies: Automated classification of pediatric sarcoma using digital histopathology. 2025 AACR Annual Meeting. Abstract 8. Presented April 28, 2025.
2. Mukashyaka P, Sheridan TB, Foroughi Pour A, et al: SAMPLER: Unsupervised representations for rapid analysis of whole slide tissue images. eBioMedicine 99:104908, 2024.
