Father-son team brings AI into hematopathology with clinical intent
Date: November 10, 2025
USF Distinguished University Professor Dmitry Goldgof and his son, Dr. Gregory Goldgof of Memorial Sloan Kettering Cancer Center (MSK), have joined forces to apply computer vision to bone marrow smears. Their latest study in Science Translational Medicine details DeepHeme, a deep learning ensemble built to identify and classify cells in blood and bone marrow - a task that typically demands hours of manual review by pathologists.
The goal is straightforward: improve diagnostic speed and consistency in hematologic diseases while keeping specialists in control. The MSK-led work shows that deep learning can quickly and accurately classify cells relevant to diagnosis, helping clinicians deliver timely, data-backed decisions for patients.
The collaboration
Dmitry Goldgof has spent more than three decades advancing algorithms and AI. Gregory Goldgof runs a translational AI lab at MSK, directing Artificial Intelligence for Hematopathology and bridging computer science with clinical practice. Together, they've built a workflow that moves AI from concept to clinic - not as a replacement for experts, but as a force multiplier.
"This is the kind of problem where computer vision and medicine come together," said Dmitry. "It's exciting to see how my expertise in algorithms and AI can complement Gregory's clinical focus."
For Gregory, teaming up was intentional. "I chose to focus on medical image analysis partly because it gave me an opportunity to work with him," he said. "We share the same passion, but my work is centered on developing clinical products for patients, while his is focused on the science of the algorithms. Together, that creates something stronger."
Why this matters for labs and patients
Diagnosing blood cancers often starts with bone marrow smears. These slides include a mix of developing red and white blood cells and platelets. Today, pathologists count and classify hundreds of cells by hand to build a differential and reach a diagnosis.
DeepHeme targets the bottleneck. By automating much of the cell detection and classification workload, it can shorten time-to-result and support more consistent, standardized reports across cases and operators.
- Speed: Automates core steps of the differential that slow down workflows.
- Reproducibility: Applies the same decision process across similar images, reducing variability.
- Scalability: Frees specialists to focus on edge cases and integrative interpretation.
What's in the approach
DeepHeme is presented as a "high-performance, generalizable deep ensemble" for bone marrow morphometry and hematologic diagnosis. In practice, that means multiple models cooperate to detect and classify cells, aggregate results, and produce outputs structured for clinical review.
Details such as training data scale, annotation protocols, and external validations are covered in the paper. The key takeaway for practitioners: the system is built for generalization and clinical fit, not just benchmark scores.
From research to deployment
The next step is deployment at MSK. As a global leader in blood cancer care, MSK provides a high-signal environment to test real-world performance: integration with lab information systems, human-in-the-loop review, and feedback cycles that refine the model under clinical supervision.
This aligns with precision oncology principles - consistent measurements, reliable differentials, and faster turnaround - while preserving expert oversight where it matters most.
How the partnership started
The collaboration began informally during the pandemic. After launching his lab, Gregory invited Dmitry into virtual meetings to advise students and postdocs across multiple projects. Over time, that cadence evolved into a practical research pipeline that connects algorithm development with clinical delivery.
For Dmitry, who has helped build USF's reputation in AI research over 35 years, seeing the work move into patient care is personal: "To be able to work together, it's very special."
Practical guidance for teams building clinical AI
- Data governance: Lock down de-identification, access controls, and audit trails. Set clear policies for model training and secondary use.
- Annotation strategy: Define cell classes, consensus rules, and adjudication early. Measure inter-rater agreement and use it to set targets.
- Generalization: Validate across stains, scanners, and sites. Plan for domain shift and include prospective monitoring.
- Human-in-the-loop: Keep final decisions with pathologists. Provide uncertainty flags, quality metrics, and fast review tools.
- Integration: Fit into LIS/EMR workflows without adding clicks. Prioritize clear summaries, diffs vs. prior cases, and exportable reports.
- Regulatory and QA: Document versioning, performance claims, and change control. Treat model updates like assay changes.
- Post-market monitoring: Track drift, failure modes, and equity metrics. Close the loop with periodic re-calibration and re-training.
Citation and further reading
DeepHeme: Shenghuan Sun et al., "DeepHeme, a high-performance, generalizable deep ensemble for bone marrow morphometry and hematologic diagnosis," Science Translational Medicine (2025). DOI: 10.1126/scitranslmed.adq2162
Background on blood cancers: NCI: Leukemia overview
Bottom line: Pair deep learning with pathologist expertise, validate across settings, and integrate tightly with lab workflows. That's how AI becomes useful in hematopathology - and how efforts like DeepHeme move from paper to practice.
Your membership also unlocks:
