AttriGen: Dual-Model Blood Cell Annotation

• AttriGen is a dual-model framework that automates extraction of a clinically aligned 12-attribute profile from blood cell microscopy images.
• The system mirrors a pathologist’s workflow by first classifying cell type (8-way) and then predicting 11 morphological attributes using CNN and transformer-based models.
• It achieves near-human performance with high accuracy metrics, significantly reducing annotation time and enabling scalable, interpretable analysis.

AttriGen is a dual-model framework for automated, fine-grained multi-attribute annotation in blood cell microscopy, designed to produce a clinically aligned 12-attribute profile per image: one cell type plus eleven morphological attributes. It was introduced to address the gap between widely studied single-label white blood cell recognition and the multi-attribute characterization used in microscopic review, where clinical reasoning depends on traits such as nuclear-to-cytoplasmic ratio, chromatin density, granularity, cytoplasm texture, and vacuolation rather than cell type alone (Houmaidi et al., 30 Sep 2025).

1. Clinical problem and conceptual scope

AttriGen is motivated by the fact that manual microscopic review of peripheral blood smears is slow, expert-dependent, and increasingly constrained by workforce shortages. In this setting, deep learning had already achieved high accuracy in coarse white blood cell type classification, but clinically relevant morphology remained comparatively underrepresented. The core problem is that morphological assessment is intrinsically multi-label per image and involves interdependent attributes, making comprehensive annotation substantially more expensive than single-type labeling (Houmaidi et al., 30 Sep 2025).

The framework is explicitly structured to mirror a pathologist’s workflow: first identify the cell type, then read morphology, and finally fuse the outputs into a richer representation. This design positions AttriGen not only as a recognition system but also as an annotation-expansion pipeline for large unlabeled corpora. The emphasis is therefore on accurate, interpretable multi-attribute prediction and on reducing the labor cost of multi-attribute dataset construction.

A central feature of the system is the distinction between coarse semantic categorization and fine-grained morphological profiling. In AttriGen, cell type is treated as an 8-way classification problem, whereas morphology is treated as an 11-attribute multi-label prediction problem. The resulting representation is more clinically aligned than traditional white blood cell classifiers because it exposes the features that pathologists use to reason about disease.

2. Datasets and attribute schema

AttriGen uses two complementary datasets: the Peripheral Blood Cell dataset for cell type recognition and the WBC Attribute Dataset for morphological annotation (Houmaidi et al., 30 Sep 2025).

Dataset	Content	Key reported properties
Peripheral Blood Cell (PBC)	Single-cell microscopy images for cell type classification	17,092 JPG images; CellaVision DM96; healthy donors; May–Grünwald–Giemsa stain; native resolution 360×363 px
WBC Attribute Dataset (WBCAtt)	Microscopy images with morphological attribute labels	10,298 images; 11 morphological attributes; annotated by biomedical students then reviewed by research scientists; pathologist involvement to resolve ambiguities

The PBC dataset contains eight classes with the following counts: neutrophils (3,329), eosinophils (3,117), basophils (1,218), lymphocytes (1,214), monocytes (1,420), immature granulocytes (2,895), erythroblasts (1,551), and platelets (2,348). Preprocessing resizes images to $224 \times 224$, normalizes pixel values to $[0,1]$, and converts labels to numeric with one-hot encoding for multi-class training. The reported split is 80% train (13,673), 10% validation (1,710), and 10% test (1,709), with balanced representation. Data augmentation uses shear=0.3 and zoom=0.3 (Houmaidi et al., 30 Sep 2025).

WBCAtt contains 11 morphological attributes: cell size, cell shape, nucleus shape, nuclear-to-cytoplasmic ratio, chromatin density, cytoplasm texture, cytoplasm color, cytoplasm vacuole, granularity, granule type, and granule color. The dataset is explicitly multi-label per image. Its preprocessing pipeline loads CSV files with image paths and attributes, encodes attributes to numeric values via alphabetically sorted bijective mappings, resizes to $224 \times 224$, normalizes intensities, converts samples to tensors, and serializes sample dictionaries. The split is approximately 60% train (6,179), 10% validation (1,030), and 30% test (3,099). ViT-based models use random horizontal flip, resize to 256 px, and center crop to 224×224; CNN-based models use random crop 224×224 and random horizontal flip.

The attribute definitions are conceptual rather than narrowly ontological. For example, cell size is described as relative cellular scale, cell shape as contour geometry, nucleus shape as nuclear morphology, and N:C ratio as the area ratio of nucleus to cytoplasm. Granularity, granule type, and granule color separate the presence, morphology, and hue of cytoplasmic granules. The annotation process required that each image be seen by at least two annotators, with pathologist involvement to resolve ambiguities, but inter-annotator agreement statistics are not reported.

3. Architecture and training regime

AttriGen uses a dual-model setup rather than a joint multitask architecture (Houmaidi et al., 30 Sep 2025). The cell type classifier is a VGG16 model pre-trained on ImageNet and fine-tuned on PBC for 8-way classification. Its output is an 8-class softmax. The morphological branch is a separate multi-attribute classifier trained on WBCAtt; it evaluates five model variants: ViT-B/16, Swin-S, DeiT-B, DeiT-S, and VGG16. The attribute head uses 11 sigmoid outputs, one for each attribute.

This separation is methodologically important. The two models are trained independently on their respective datasets, and no joint training or parameter sharing is reported. During inference, their predictions are fused by concatenation into a 12-attribute profile. No algorithmic cross-talk or knowledge transfer is described beyond this fusion step. This makes AttriGen a modular pipeline rather than a unified shared-backbone multitask learner.

Among the attribute models, the paper emphasizes the contrast between CNNs and transformers. VGG16 serves as the CNN baseline for attribute recognition, whereas ViT-B/16 uses 16×16 patches and a transformer encoder stack, Swin-S uses shifted-window hierarchical attention, and DeiT-B/S are data-efficient Vision Transformers with distillation strategies. The reported architectural conclusion is that transformer-based models outperform the CNN baseline on WBCAtt.

The training configuration is only partially specified. Both tasks use input resolution $224 \times 224$ and normalize pixels to $[0,1]$. Optimizer, batch size, epoch count, learning-rate schedule, weight decay, compute hardware, and training time are not disclosed. The paper also does not report decision thresholds for attribute inference, class-imbalance handling, calibration metrics, or regularization terms. It states that the loss formulations are not explicitly documented, although the task structure naturally implies standard cross-entropy for cell type prediction and binary cross-entropy for the 11-way multi-label attribute head.

4. Evaluation protocol and empirical results

AttriGen reports accuracy, precision, recall, and F1-score per attribute head and per task. For multi-attribute evaluation it defines a global average accuracy over attributes as

$$\mathrm{Acc}_{\mathrm{global}} = \frac{1}{N}\sum_{i=1}^{N} \mathrm{Acc}_i$$

where $N$ is the number of attribute heads (Houmaidi et al., 30 Sep 2025).

Model/task	Metric	Reported result
VGG16 on PBC	Test accuracy	98.83%
Swin-S on WBCAtt	Global average accuracy	94.62%
ViT-B/16 on WBCAtt	Global average accuracy	94.35%
DeiT-B on WBCAtt	Global average accuracy	94.31%
DeiT-S on WBCAtt	Global average accuracy	94.19%
VGG16 on WBCAtt	Global average accuracy	91.42%
Prior baseline (MAP)	Global average accuracy	88.65%

The PBC type classifier reaches 98.83% test accuracy. The paper notes that several external works report higher accuracies up to 99.91%, but states that maximizing the cell-type benchmark was not its primary objective. The more central result is on WBCAtt, where Swin-S establishes a new benchmark with 94.62% global average accuracy (Houmaidi et al., 30 Sep 2025).

The Swin-S attribute-wise breakdown shows marked heterogeneity across morphological traits. N:C Ratio reaches 98.61% accuracy, Cytoplasm Vacuole 97.58%, Cytoplasm Texture 96.55%, Cytoplasm Color 95.58%, Chromatin Density 94.93%, and Cell Shape 94.29%. By contrast, Cell Size is 84.03% and Nucleus Shape 80.51%, making shape-linked attributes the most difficult among the reported categories. Granule-related cues are especially strong: Granule Type 99.61%, Granule Color 99.29%, and Granularity 99.81%. The reported error analysis attributes this pattern to transformers’ ability to combine global and local context, which appears particularly effective for granule-related morphology and N:C ratio.

The paper also compares AttriGen with a human expert benchmark. It reports a 1.48% gap between the best model’s 94.62% GAA and an expert benchmark at 96.1%, and interprets this as near-human performance with substantial efficiency gains. Confusion matrices and calibration metrics are not reported. The paper additionally notes that prior methods such as P4BC and ConceptCLIP use different metrics—ACAcc 63.6% and AUC 99.5%, respectively—thereby motivating its use of GAA as a metric with equal weighting across attributes.

5. Interpretability and annotation workflow

Interpretability in AttriGen is qualitative rather than formalized through dedicated explainability metrics (Houmaidi et al., 30 Sep 2025). The study applies Grad-CAM to the Swin-S attribute classifier and reports that heatmaps typically illuminate nucleus-centric regions for N:C ratio and chromatin density, cytoplasmic regions for texture, color, and vacuole prediction, and granule clusters for granularity, granule type, and granule color. These visualizations are presented as evidence that model attention aligns with clinically relevant morphology.

This interpretability layer is central to the framework’s stated purpose. AttriGen is not described merely as a predictive model but as a means of making downstream models more interpretable by expanding datasets with explicit morphological labels. Because the outputs are decomposed into clinically meaningful attributes, the pipeline yields a richer annotation structure than type-only classifiers. A plausible implication is that this representation is more suitable for expert review, auditing, and curation than a single categorical prediction, although the paper limits itself to qualitative interpretability claims.

The annotation pipeline itself is framed as a bootstrapping procedure. A small expert-labeled multi-attribute dataset, WBCAtt, is used to train and validate the attribute recognizer until near-human accuracy is obtained; the trained attribute model and the cell-type classifier are then applied to larger unlabeled corpora. For the PBC dataset, the paper states that annotating the remaining 6,784 unlabeled cells at a rate of 20 ms per cell takes approximately 2.26 minutes, in contrast to weeks for exhaustive human annotation with multi-attribute review. The significance of AttriGen therefore lies not only in its benchmark performance but also in its function as a scalable annotation generator.

6. Limitations, reproducibility, and broader implications

The study identifies several limitations that constrain current generalization claims (Houmaidi et al., 30 Sep 2025). PBC samples healthy donors, uses May–Grünwald–Giemsa staining, and originates from a specific device, the CellaVision DM96; this may limit transfer to other laboratories, staining protocols, devices, or pathological conditions. WBCAtt labels, although curated, remain potentially subjective, and inter-annotator agreement is not quantified. The paper also reports no explicit tests under domain shift, staining variation, different microscopes, or small-sample regimes.

A second limitation concerns experimental disclosure. Although the paper provides dataset identities, train/validation/test splits, input resolution, preprocessing, augmentations, model families, and label encodings, it does not specify optimizer, learning rate, batch size, epochs, weight decay, training time, hardware, or code/checkpoint availability. Other detailed architectural hyperparameters are also omitted beyond standard model names and the fact that ViT-B/16 uses 16×16 patches. As a result, the work is reproducible at the level of task definition and coarse model setup, but not at the level of a full training recipe.

The proposed extensions are correspondingly pragmatic. The paper suggests expanding the framework to pathological and abnormal morphologies important for leukemia diagnosis, incorporating active learning to focus expert review, and improving explainability methods to better link attributes to image regions. It also presents AttriGen as a conceptual template for other computer vision tasks where fine-grained multi-attribute labeling is clinically or scientifically necessary but prohibitively expensive to scale manually.

In that sense, AttriGen occupies a distinct position within blood-cell image analysis. It does not primarily seek to maximize conventional white blood cell classification benchmarks. Instead, it reframes automated hematology microscopy around the clinically aligned expansion of attribute labels, using a CNN type classifier and a transformer attribute classifier to convert sparse expert supervision into large-scale structured annotations. The main reported contribution is therefore the operationalization of multi-attribute blood-cell annotation as a scalable, interpretable computer-vision workflow rather than a narrow single-task classifier (Houmaidi et al., 30 Sep 2025).