A new study published in Nature Machine Intelligence describes a generative artificial intelligence model, CytoDiffusion, designed to classify blood cell images while also identifying morphologies that may warrant specialist review.
Blood film review remains a core part of hematology diagnostics, but consistent classification of leukocyte morphology is difficult to automate at scale. Cell appearance varies with staining, slide preparation, imaging systems, and disease state. Meanwhile, rare or atypical cells can fall outside the categories that algorithms are trained to recognize.
Most morphology classifiers used in digital pathology are “discriminative” models, trained to separate predefined cell types. CytoDiffusion takes a different approach. Built on a diffusion-based generative framework, it learns the visual range of each blood cell class, rather than relying only on decision boundaries. The authors argue this may improve performance under real-world conditions, particularly when slides or image sources differ from the training data, and when abnormal cells appear that the model has not previously encountered.
The researchers trained the model using a dataset of 32,619 blood cell images. To assess how realistic the generated images were, 10 expert hematologists were asked to distinguish real images from synthetic ones. Their accuracy was close to chance (0.523), suggesting the synthetic morphology was difficult to reliably separate from real microscopy images. When the same experts were asked to classify the synthetic cells by type, their labels closely matched the intended classes (agreement 0.986), indicating that key class-specific features were preserved.
The authors then tested CytoDiffusion against established blood cell classification benchmarks and compared it with commonly used discriminative models, including vision transformers and EfficientNet variants. Across multiple datasets, CytoDiffusion performed comparably to or better than these methods, while also offering additional capabilities relevant to diagnostic workflows – most notably uncertainty handling and anomaly detection.
CytoDiffusion was also noted for its ability to flag “out-of-distribution” cells: morphologies that do not match the expected patterns of common leukocyte classes. In experiments where important abnormal cells were excluded during training, CytoDiffusion performed well at identifying these atypical images. In a dataset designed to test abnormal cell detection using blasts as the held-out class, the model achieved a sensitivity of 0.91 and specificity of 0.96, compared with substantially lower sensitivity for a vision transformer comparator (0.28).
For diagnostic laboratories, this distinction matters. In practice, a morphology tool is most useful when it can support triage – handling routine cells efficiently while escalating uncertain or abnormal cases for expert confirmation. The authors also describe visualization tools, including “counterfactual heat maps,” intended to show which image regions influenced classification decisions, potentially supporting interpretability in review workflows.
The study also reports a practical limitation: the approach is more computationally intensive than conventional classifiers, with mean classification time reported at approximately 1.8 seconds per image, which may require optimization before routine high-throughput implementation.
Overall, the findings highlight a potential shift in digital morphology from single-label classification toward systems that can quantify uncertainty and better recognize rare or unfamiliar cell patterns – an important requirement for safe deployment in hematology diagnostics.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
