A few months ago, we announced the participation of Artikode Intelligence in DEISA, a project funded under the NextGenerationEU program aimed at developing artificial intelligence tools for advanced diagnosis in digital pathology. With the project now coming to an end, we share the results of this initiative, whose objective has been to create an automated system capable of supporting the differential diagnosis of Ewing Sarcoma versus tumors of the Ewing-like family.
Accurate diagnosis of these tumors is particularly complex. Ewing Sarcoma presents histopathological features that overlap with those of other small round cell sarcomas, and Ewing-like tumors are also extremely rare and share similar morphological patterns. These similarities make differential diagnosis a clinical challenge even for experienced specialists and highlight the need for tools that help analyze histological images in a more objective and reproducible way.
To address this problem, the DEISA consortium developed a deep learning–based framework capable of analyzing whole histological images (Whole Slide Images, WSI) and performing automatic subtyping. For model training, a multicenter dataset was built consisting of 64 Ewing Sarcoma samples and 143 Ewing-like tumors, including variants such as CIC, BCOR, NFATC2, and PATZ1.
On this dataset, a Multiple Instance Learning (MIL) architecture was implemented, dividing the diagnostic process into two stages: first, distinguishing Ewing Sarcoma from the Ewing-like family, and then performing multiclass classification to identify the specific subtype.
The results obtained demonstrate the potential of artificial intelligence as a diagnostic support tool. In the first stage, the model differentiated Ewing Sarcoma from the Ewing-like family with an accuracy of 85.5% and a sensitivity of 71.8%. In the detailed classification stage, it achieved an accuracy of 71.6%, with particularly high sensitivity in detecting the CIC variant. Overall, the complete diagnostic workflow reached a global accuracy of 67.6%, confirming the feasibility of this approach in a real digital pathology environment.
Throughout the project, the PixNormous platform developed by Artikode has been the technological component responsible for hosting the algorithms, managing the visual datasets, and enabling scalable execution of predictions on large images. Integrating the models into a stable platform has made it possible to work with complex data in a reproducible way, facilitating the transition from research to potential real-world application.
From a rigorous scientific perspective, these results do not represent a final solution, but a necessary step forward. DEISA has made it possible to create an automated framework that lays the groundwork for the real integration of artificial intelligence in the diagnosis of Ewing Sarcoma and in the study of rare tumors. At Artikode, we are proud to have contributed our technology to a project that opens the door to new decision-support tools for specialists in the field of digital pathology.






