AI in Leukemia Microscopy: A Research Review

Hybrid image models: Meta-Conformer-XAI

Jammal's Scientific Reports paper combines convolutional and transformer components with attention and interpretability methods. The author reports accuracy of 99.24% and 96.36% on two evaluated datasets. These figures belong to the study's experiments; they are not CellSight results or evidence of universal clinical performance. ^[1]

The methodological question is how local visual features and broader context contribute under the specified evaluation. For practical assessment, readers should inspect the patient partitions, class balance and preprocessing rather than relying on the accuracy figure alone.

Detection within images: SGLNet

Mei and colleagues introduce a spatially guided network and a dataset of 1,794 bone marrow microscopy images. The published abstract reports mean average precision of 95.9% and 98.6% for the evaluated ALL and CLL detection tasks. Mean average precision is a detection measure and should not be relabelled as patient-level diagnostic accuracy. ^[2]

Throughput comparisons need the unit, hardware and complete processing pipeline before they can be interpreted. A rate measured on selected images should not be assumed to describe the processing time for complete slides.

Combining cell information: CAREMIL

Singi and colleagues combine DeepHeme image encoding with attention-based multiple instance learning to aggregate information across cells. The published study reports AUROCs of 0.999 for AML, 0.891 for MDS and 0.945 for HCL in its evaluated tasks. MDS is included here as a distinct hematologic condition, not as a leukemia subtype. ^[3]

This work illustrates the difference between assigning a label to an individual cell and integrating information at case level. An attention view can help a reader inspect influential cells; it should not automatically be treated as a complete explanation of a model's reasoning.

Comparing established architectures

Ashikuzzaman and colleagues compare fourteen pretrained convolutional architectures using a 10,700-image dataset and include explainability analysis. This provides a documented example of comparative model research. It does not establish that an architecture with a favourable result in one experiment will perform best in another laboratory. ^[4]

For a useful comparison, the data, split, processing and metric must remain visible. A large image count should also be distinguished from the number and diversity of independent patients represented.

What these studies do and do not establish

Our interpretation is that the studies offer different methodological ideas: combining local and contextual features, detecting cells within fields, aggregating cell evidence and comparing established model families.

Their reported measures cannot be placed in one league table. Accuracy, detection mean average precision and AUROC describe different properties. External validation, calibration and workflow evaluation remain questions to examine within each study; none should be inferred solely from a favourable number.

This is a selected narrative review, not an exhaustive systematic review. It focuses on the methods and reported evaluations of the four cited studies.

Relevance to CellSight

For CellSight, this literature informs how to ask better research questions and document technical choices. It does not establish that these architectures are implemented in CellSight or that their reported results apply to the product. CellSight's working first version is a research platform and is not for diagnosis or clinical decision-making.