SYNTHEMA, represented by consortium members Humanitas Research Hospital and Alma Mater Studiorum – Università di Bologna, made a significant contribution at the 65th ASH Annual Meeting and Exposition in San Diego. This premier event in classical and malignant hematology, held from December 9-12, 2023, was a blend of in-person and online sessions. The meeting included a variety of scientific and educational programs, specialized workshops, symposia, and sessions focusing on diverse aspects of hematology.

SYNTHEMA’s representatives showcased groundbreaking work in the field of hematology. They presented their latest research on generating histopathological synthetic images using stable diffusion models and variational autoencoders, particularly focusing on Acute Myeloid Leukemia (AML). This innovative approach utilized data from Tazi et al’s paper, offering new insights into AML classification and risk stratification.

Additionally, the team validated their dataset by extracting morphological and texture features from synthetic images, comparing them to real images to predict critical statistics like Overall Survival and Progression-Free Survival. This methodology extended to sequencing synthetic data, where they employed Dirichlet clustering and Markov models for state transitions, benchmarking their results against actual data.

SYNTHEMA also highlighted their work on Sickle Cell Disease (SCD), employing a dual approach on available Brain MRI data of SCD patients. They generated synthetic images and features, comparing these with real data to ensure accuracy and reliability. Their presentations and contributions to the meeting emphasized the role of advanced computational techniques in enhancing our understanding and treatment of hematological disorders, marking a significant step forward in medical research and patient care.

SYNTHEMA consortium representatives at the ASH23

Publication: Clinical Text Reports to Stratify Patients Affected with Myeloid Neoplasms Using Natural Language Processing

Authors: Gianluca Asti, Elisabetta Sauta, Nico Curti, Gianluca Carlini, Lorenzo Dall’Olio, Luca Lanino, Giulia Maggioni, Alessia Campagna, Marta Ubezio, Antonio Russo, Gabriele Todisco, Cristina Astrid Tentori, Pierandrea Morandini, Marilena Bicchieri, Maria Chiara Grondelli, Matteo Zampini, Erica Travaglino, Victor Savevski, Nicolas Riccardo Derus, Daniele Dall’Olio, Claudia Sala, Lin-Pierre Zhao, Armando Santoro, Shahram Kordasti, Valeria Santini, Anne Sophie Kubasch, Uwe Platzbecker, Maria Diez-Campelo, Pierre Fenaux, Amer M. Zeidan, Torsten Haferlach, Gastone Castellani, Matteo Giovanni Della Porta, Saverio D’Amico

Published in: 65th ASH Meeting in San Diego |803.EMERGING TOOLS, TECHNIQUES AND ARTIFICIAL INTELLIGENCE IN HEMATOLOGY

Abstract:  This study, spearheaded by GenoMed4All and Synthema EU consortia, showcases the development of HematoBERT, an AI language model tailored for the hematology domain. Utilizing the BERT framework, HematoBERT was fine-tuned with hematological clinical reports, focusing on diseases like myeloproliferative neoplasms, myelodysplastic syndrome, and acute myeloid leukemia. This model successfully clustered patients based on their clinical reports, revealing significant insights into patient stratification and prediction of clinical outcomes. HematoBERT demonstrated superior performance in understanding contexts and correlations compared to pre-trained non-contextualized models, highlighting its potential in enhancing personalized medicine through the integration of clinical and genomic data.

Read the full paper here

Publication: Synthetic Histopathological Images Generation with Artificial Intelligence to Accelerate Research and Improve Clinical Outcomes in Hematology

Authors: Gianluca Asti, Saverio D’Amico, Nico Curti, Gianluca Carlini, Elisabetta Sauta, Nicolas Riccardo Derus, Daniele Dall’Olio, Claudia Sala, Lorenzo Dall’Olio, Luca Lanino, Giulia Maggioni, Alessia Campagna, Marta Ubezio, Antonio Russo, Gabriele Todisco, Cristina Astrid Tentori, Pierandrea Morandini, Marilena Bicchieri, Maria Chiara Grondelli, Matteo Zampini, Victor Savevski, Armando Santoro, Shahram Kordasti, Valeria Santini, Anne Sophie Kubasch, Uwe Platzbecker, Maria Diez-Campelo, Pierre Fenaux, Lin-Pierre Zhao, Amer M. Zeidan, Torsten Haferlach, Gastone Castellani, Matteo Giovanni Della Porta

Published in: 65th ASH Meeting in San Diego |803.EMERGING TOOLS, TECHNIQUES AND ARTIFICIAL INTELLIGENCE IN HEMATOLOGY

Abstract:  In this groundbreaking paper by GenoMed4All and Synthema EU consortia, the focus was on utilizing AI to generate synthetic hematological images from textual data, addressing the challenges in collecting multimodal data for rare and complex hematological malignancies. The initiative involved applying the Stable Diffusion generative model, fine-tuned with hematological data, to create Hematoxylin and Eosin images for myeloid neoplasm patients. This was complemented by HematoBERT, a domain-specific language model encoding textual inputs to condition image generation. The study also developed a Synthetic Images Validation Framework (SIVF) to evaluate the utility and fidelity of these synthetic images. The results showed that synthetic data significantly enhanced disease classification and prediction of survival probabilities, with synthetic augmentation improving model performances by over 10%. This innovative approach demonstrated the potential of AI-generated images in replicating real-world morphological features, thereby boosting precision medicine research in hematology through effective data augmentation and simplified data sharing.

Read the full paper here

Leave a Reply

Your email address will not be published. Required fields are marked *