White blood cells are the immune system’s sentinels, and their relative proportions in a blood smear are among the most diagnostically powerful numbers in medicine. A surge in neutrophils can signal a bacterial infection; a rise in lymphocytes often points to viral illness; elevated eosinophils betray allergies or parasitic invasion; and blasts—immature cells that should never appear in peripheral blood—are the alarm bell of acute leukemia. For decades, producing this picture has required a trained hematologist to scan a stained smear under a microscope and manually tally hundreds of cells, a process that is slow, expensive and vulnerable to fatigue and inter-observer disagreement. Automated hematology analyzers can count cells by the thousand, but when their results flag an abnormality, laboratories still fall back on visual microscopy review, the accepted gold standard for morphological confirmation. Deep learning promises to close that gap: a system that could triage slides automatically, standardize readings across hospitals and alert clinicians to suspicious cells within seconds would fundamentally change the economics and reliability of hematological diagnostics.

To bring order to this fast-moving literature, the team conducted a systematic literature review following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses, the internationally recognized PRISMA protocol that governs how studies are searched, screened and included. Rather than simply cataloguing papers, the authors evaluated each study against an analytical framework of 18 technical dimensions, covering everything from dataset selection and preprocessing to model architecture, training strategy and reported performance. The review was organized around three research questions: which deep learning models dominate the field, how image preprocessing and segmentation shape classification performance, and what limitations and future directions define the state of the art. Studies were selected and analyzed on the basis of relevance, methodological rigor and contribution to the field. The work was funded in part by the European Union’s REFRESH—Research Excellence For Region Sustainability and High-tech Industries project and by the Ministry of Education of Czechia, and was led by corresponding author Rene Jaros of VSB–Technical University of Ostrava.

The headline statistic is unambiguous: convolutional neural networks appear in 81.6 percent of the reviewed studies, making them the undisputed backbone of white blood cell image analysis. CNNs are engineered to exploit the spatial structure of images. Instead of treating a micrograph as a flat list of pixels, they slide small filters—known as convolution kernels—across the image, each filter learning to fire in response to a specific visual pattern such as an edge, a curve, a granule or the lobed contour of a nucleus. Stacked in layers, these filters build a hierarchy of features: early layers detect simple lines and color transitions, while deeper layers assemble them into the complex shapes that distinguish a segmented neutrophil from an eosinophil or a monocyte. Crucially, such networks are rarely trained from scratch. The standard approach is transfer learning: the network is initialized with weights learned from ImageNet, a corpus of more than a million everyday photographs, and then fine-tuned on much smaller blood cell datasets, dramatically reducing the amount of scarce medical data required to reach high accuracy.

Within the CNN family, three architectures tower over the rest. ResNet, used in 32.9 percent of the studies, introduced a deceptively simple trick that transformed deep learning: residual connections, shortcut pathways that let each layer learn only the difference—a residual—between its input and the desired output. These skip connections allow gradients to flow unimpeded through dozens or even hundreds of layers during training, curing the vanishing-gradient problem that had previously made very deep networks untrainable. The Visual Geometry Group network, or VGG, chosen in 18.4 percent of studies, takes the opposite philosophical route: it relies on a plain, disciplined stack of small three-by-three convolution filters, repeated layer after layer, whose simplicity makes it easy to retrain and adapt to new cell datasets. EfficientNet, the youngest of the trio at 14.5 percent, takes a more surgical approach. Rather than scaling the network’s depth, width or input resolution independently, it scales all three together according to a compound coefficient, squeezing state-of-the-art accuracy out of a fraction of the parameters—an efficiency that matters when models must eventually run on laboratory workstations rather than data-center GPUs.

Convolutional networks are not the only players. Hybrid approaches, which combine a CNN with a second machine-learning engine, account for 11.2 percent of the literature, and their logic is straightforward: let each component do what it does best. In one common pairing, a CNN first digests a cell image and outputs a compact vector of learned features; a support vector machine then draws the final class boundaries in that high-dimensional feature space, exploiting its strength in maximizing the margin between categories. Other teams pair CNNs with long short-term memory networks—recurrent layers designed to remember information across sequences—an appealing option when cells arrive in streams from automated scanners and temporal context can inform a decision. Such hybrids persist because they pair complementary strengths: the CNN supplies rich visual representations while the second model refines the decision, sometimes at lower computational cost.

Before any network sees a cell, another family of techniques quietly shapes the outcome. Preprocessing methods, including histogram equalization, normalization and contrast enhancement, were employed in 68 percent of the reviewed studies, and for good reason. Blood smears are stained with dyes such as Giemsa or Wright’s stain, and the resulting colors vary with staining time, reagent batch, microscope illumination and the scanner or camera used to digitize the slide. Left uncorrected, this variability can cause a model trained in one hospital to fail in another. Histogram equalization redistributes the intensity values of an image so that they span the full dynamic range more evenly, stretching faint contrasts into visible ones; normalization rescales pixel values to a standard range, stabilizing the arithmetic of training; and contrast enhancement sharpens the boundary between nucleus, cytoplasm and background. Together these steps act as a common language that lets images from different laboratories, microscopes and staining protocols feed into the same model without confusing it—a quiet but decisive contributor to the accuracy figures reported across the field.

Segmentation, used in 54 percent of the studies, goes a step further by teaching the pipeline where each cell actually is. Two techniques recur most often. U-Net, an architecture originally invented for biomedical image segmentation, uses an encoder-decoder design: a contracting path that compresses the image into abstract feature maps, followed by an expanding path that reconstructs a full-resolution, pixel-by-pixel mask, with skip connections that shuttle fine spatial detail from the encoder directly to the decoder. The result is a precise outline of the nucleus and cytoplasm of every cell. The watershed algorithm works differently, treating the grayscale image as a topographic landscape in which dark nuclei form basins; flooding those basins from their minima partitions the image so that touching or overlapping cells are separated along ridge lines. Segmentation matters because morphology is diagnosis: nuclear shape, chromatin texture, granule distribution and the nucleus-to-cytoplasm ratio are exactly the features hematologists weigh, and isolating the cell from red blood cells, platelets and background debris lets a classifier focus on what counts.

The single most eye-catching number in the review is 99.83 percent: the highest classification accuracy reported anywhere in the surveyed literature, achieved with EfficientNet on the BCCD dataset, a widely used public collection of annotated blood cell micrographs. Such figures fuel the viral claim that AI now outperforms human microscopists, but the authors are careful to place them in context. Benchmark datasets are typically small, curated and carefully balanced, whereas real-world smears present overlapping cells, staining artifacts, rare cell types and scanner-to-scanner variation that no curated test set fully captures. A model that reaches near-perfect accuracy on a benchmark may lose several percentage points—or fail in unexpected ways—when confronted with data from an unfamiliar laboratory. This is why the review insists that reported accuracy must be read alongside dataset provenance, class balance and validation design. It also introduces its 18-dimension analytical framework precisely to make such comparisons honest, giving researchers a common yardstick for benchmarking existing studies rather than a patchwork of incomparable accuracy claims.

What comes next, the authors argue, is less about chasing ever-higher accuracy on familiar datasets and more about making deep learning fit for the clinic. That means models that are robust to domain shift, explainable enough for a hematologist to audit—so the system can show which visual features drove a decision—and deployable within the real workflow of a hematology laboratory. It also means larger, more diverse and more carefully annotated datasets spanning institutions, staining protocols and disease spectra, because the diversity of training data, not architectural novelty alone, will determine whether laboratory-to-laboratory generalization is achieved. The review’s structured framework is intended to guide exactly this development, steering the field toward diagnostic tools that are not merely impressive in papers but trustworthy at the bench. With white blood cell analysis sitting at the intersection of oncology, immunology and infectious disease, the stakes extend far beyond the laboratory: a dependable AI pipeline for reading blood smears could bring specialist-grade hematological assessment to hospitals and clinics that have never had a hematologist on call.