“We’re transforming a century-old tool into a quantitative platform for immune monitoring”
At the 52nd Annual Meeting of the European Society for Blood and Marrow Transplantation (EBMT) in Madrid, Dr Roni Shouval (Memorial Sloan Kettering Cancer Center (MSKCC), New York, NY, USA) presented his plenary abstract: a novel approach to one of the field’s persistent challenges: scalable, real-time immune monitoring following CAR-T therapy. By applying high-throughput deep learning to routine peripheral blood smears, his team is transforming a century-old diagnostic tool into a quantitative platform capable of capturing immune dynamics longitudinally. In this interview, Dr Shouval discusses the clinical rationale, technological framework, and future implications of this emerging methodology.
A key gap in CAR-T care is the lack of a scalable, real-time and routine method to monitor immune responses. Existing tools – such as flow cytometry or molecular assays – offer high biological resolution, but they are resource-intensive and not always embedded in day-to-day clinical workflows. Peripheral blood smears, by contrast, are universally available. Our goal was to determine whether AI could transform this routine test into a quantitative, longitudinal immune monitoring platform that is both accessible and scalable.
Traditional assays provide detailed biological insights but are limited by cost, complexity and scalability. Our approach leverages lower-resolution image data from blood smears but applies computational methods to extract meaningful patterns. It sits somewhere in between – less granular biologically, but far more accessible and amenable to real-time monitoring. The intention is not to replace existing tools, but to complement them with a more scalable solution.
We begin with digitised peripheral blood smears, segmenting images into individual white blood cells. Expert hemopathologists then annotate lymphocyte morphologies, identifying recurring visual patterns – what we term ‘morphotypes’. Using these curated datasets, we train convolutional neural networks to recognize these morphological states. Importantly, we maintain strict separation between training and testing datasets to ensure robustness. Once trained, the model can analyse large volumes of cells and reconstruct longitudinal immune patterns following CAR-T infusion.
One key finding is that lymphocytes post-CAR-T infusion are morphologically heterogeneous. We identified reproducible cell states that evolve over time and differ depending on the CAR-T product used. Importantly, expansion of certain morphological classes correlated with improved progression-free survival.
We also observed associations between specific morphotypes and functional immune states, including CAR-T expansion. While further validation is needed, some of these morphologies may directly represent CAR-T cells, offering a potential window into treatment dynamics using routine microscopy.
We are moving from proof-of-concept toward clinical applicability. This approach builds on existing workflows, so integration is more feasible than entirely new platforms. However, prospective validation, cross-site standardisation and demonstration of added clinical value remain essential next steps.
All of the above are barriers. Models must generalise across different scanners and institutions, and regulators will require robust evidence of reproducibility and clinical utility. Equally important is clinician trust – AI must be viewed as an augmentative tool rather than a replacement for expertise. Interpretability and seamless workflow integration will be critical for adoption.
AI has the potential to shift hematology from descriptive to quantitative diagnostics. It will allow us to extract deeper insights from routine data sources – blood smears, bone marrow samples, flow cytometry and genomics. While diagnostics will likely see the greatest impact, there are also promising applications in areas such as radiology.
This work demonstrates that even a century-old tool like the peripheral blood smear can be reimagined. With AI, it can become a quantitative and potentially transformative modality for monitoring CAR-T therapy in routine clinical care.
About the European Society for Blood and Marrow Transplantation (EBMT)
The EBMT is a community of healthcare professionals, involved in clinical haematopoietic cell transplantation and cellular therapy, who share their experiences and develop co-operative studies. The EBMT has a governing body called the Board of Association and three sets of groups that channel the society’s research aims and other activities: the EBMT Working Parties, Committees, and Nurses Group, which addresses issues within the field specifically related to nursing.
This content has been developed in collaboration with the European Society for Blood and Marrow Transplantation for touchHEMATOLOGY. Views expressed are the speaker’s own and do not necessarily reflect the views of Touch Medical Media.
Disclosures: Roni Shouval is a consultant and member of the Advisory Board for AstraZeneca; he has received grant/research support and honoraria/honorarium from GSK.
Cite: From microscope to machine learning: Scaling CAR-T monitoring with AI. touchHEMATOLOGY. 7th April 2026.
Interviewer: Caroline Markham
Related content
- FDA grants orphan drug status to zavabresib in myelofibrosis
- PARADIGM challenges role of intensive chemotherapy in newly diagnosed AML
- FDA grants orphan drug status to zavabresib in myelofibrosis
SIGN UP to touchHEMATOLOGY!
Join our global community today for access to thousands of peer-reviewed articles, expert insights and learn-on-the-go education across 150+ specialties, plus concise email updates and newsletters so you never miss out.
