CP-AnemiC: Conjunctival Pallor Dataset

• CP-AnemiC is a publicly available conjunctival pallor dataset designed for non-invasive anemia detection in children, annotated with hemoglobin levels and demographic metadata.
• It supports MobileNet-based binary classification with 5-fold cross-validation and evaluates precision–performance trade-offs across FP32, FP16, INT8, and INT4 quantization.
• The dataset comprises 710 images from ten healthcare facilities in Ghana, providing a clinically anchored benchmark for mobile healthcare inference and deployment studies.

CP-AnemiC is a publicly available conjunctival pallor image dataset designed for non-invasive, image-based anemia classification in children. In "Performance Analysis of Post-Training Quantization for CNN-based Conjunctival Pallor Anemia Detection" (Romero et al., 20 Jul 2025), it serves as the primary benchmark for a MobileNet-based binary classifier and for post-training quantization experiments; the same paper notes that the dataset is also referenced in the bibliography as "Cp-anemic: A conjunctival pallor dataset and benchmark for anemia detection in children." The dataset comprises 710 conjunctival images from children aged 6–59 months and is annotated with hemoglobin levels, anemia labels, and demographic/contextual metadata, with the stated purpose of enabling machine learning and deep learning models to classify anemic versus non-anemic children.

1. Scope, provenance, and intended role

CP-AnemiC is presented as a conjunctival pallor dataset specifically constructed for anemia detection in children. Its stated goal is to support non-invasive classification using the conjunctiva as the anatomical site of interest. The paper explicitly positions the dataset as a response to limitations in earlier work, including proprietary or limited datasets, insufficient diversity, and the lack of publicly available benchmark data (Romero et al., 20 Jul 2025).

The reported collection period is January to June 2022. Images were collected from ten healthcare facilities in Ghana, and the paper describes this as providing diversity across ten regions in Ghana. The dataset is characterized as publicly available and large-scale within the paper’s framing.

This positioning matters methodologically because the paper uses CP-AnemiC not merely as a source of training images, but as a benchmark resource for clinically grounded binary classification. A plausible implication is that the dataset is intended to support both algorithmic comparison and deployment-oriented evaluation, rather than only proof-of-concept model fitting.

2. Clinical labels, metadata, and cohort composition

Each image is annotated with hemoglobin level, and the binary anemia label is derived from a threshold based on the WHO anemia criterion for children aged 6–59 months: anemic corresponds to Hb $< 11$ g/dL, and non-anemic corresponds to Hb $\geq 11$ g/dL. Additional metadata include age, gender, collection site, and remarks from laboratory assessments (Romero et al., 20 Jul 2025).

The paper reports the following quantitative summary.

Item	Value
Total images/patients	710
Age range	6–59 months
Anemic	424
Non-anemic	286
Female	306
Male	404
Mean age overall	$31.58 \pm 16.78$ months
Mean age anemic	$31.04 \pm 17.02$ months
Mean age non-anemic	$32.31 \pm 16.46$ months

The demographic summary also notes 174 female anemic cases and 132 female non-anemic cases, alongside 250 male anemic cases and 154 male non-anemic cases. The dataset includes Hb concentration analyses by age and gender, and the paper notes that anemic participants had lower Hb levels, as expected.

A common simplification is to treat CP-AnemiC as perfectly balanced. The paper’s own counts do not support that interpretation: the distribution is 424 anemic cases, or 60%, and 286 non-anemic cases, or 40%. The gender distribution is also uneven, with 306 females, or 43%, and 404 males, or 57%. The paper therefore supports the more precise description of moderate rather than severe imbalance.

3. Collection, curation, and input representation

The paper states that CP-AnemiC includes both image data and associated clinical and demographic metadata. It does not provide an extensive acquisition-standardization protocol, but it does describe the dataset as curated for benchmark use and includes visual illustrations of sample images and region-of-interest examples. The input to the learning system is a conjunctival pallor region-of-interest, although the image preprocessing pipeline is not described in detail beyond that (Romero et al., 20 Jul 2025).

The reported information on quality control is limited. The paper mentions remarks from laboratory assessments and the use of Hb thresholds for label assignment, but it does not explicitly report missing values, missing annotation rates, exclusion criteria, image quality filtering, or inter-annotator agreement. This suggests that the paper’s emphasis is on downstream model evaluation rather than on a formal dataset-audit framework.

From a dataset documentation perspective, the presence of Hb-based labels and laboratory remarks anchors the benchmark in clinically meaningful supervision. At the same time, the absence of detailed reporting on missingness and acquisition control constrains assessment of label completeness, site effects, and image-quality heterogeneity.

4. Use in the MobileNet training pipeline

In the reported experiments, CP-AnemiC is the primary training and evaluation resource for a MobileNet-based binary classifier. The model uses a pretrained MobileNet backbone initialized from ImageNet pretrained weights and is fine-tuned end-to-end on CP-AnemiC. The final layer is modified for binary classification, with output 1 denoting anemic, output 0 denoting non-anemic, and sigmoid activation (Romero et al., 20 Jul 2025).

The experimental protocol uses 5-fold cross-validation. In each fold, 4 folds are used for training and 1 fold for testing. The paper also refers to a validation step used to monitor F1 score and apply early stopping, although the exact internal allocation for validation is not described in detail. Model weights are randomly initialized at the start of each fold.

Training employs random horizontal flip, random rotation, random shifts, and random scaling for augmentation. The reported hyperparameters are batch size 32, Adam optimization, learning rate $1 \times 10^{-4}$, Binary Cross-Entropy loss, and a maximum of 150 epochs. Early stopping is triggered if there is no improvement in F1 score for 10 consecutive epochs. The model state with the best validation F1 score is saved, and the paper notes that early stopping was triggered after 26 epochs in the reported run due to limited improvement.

The dataset’s role in this pipeline is more than passive. According to the paper, CP-AnemiC supports binary anemia classification through Hb-based ground truth, provides sufficient demographic and site diversity to support cross-validation, and is suitable for evaluating lightweight MobileNet models in deployment-oriented settings.

5. Evaluation metrics and quantization formalism

The paper defines the classification metrics as follows:

$$\text{Accuracy} = \frac{\text{tp} + \text{tn}}{\text{tp} + \text{tn} + \text{fp} + \text{fn}}$$

$$\text{Precision} = \frac{\text{tp}}{\text{tp} + \text{fp}}$$

$$\text{Recall} = \frac{\text{tp}}{\text{tp} + \text{fn}}$$

$$\text{F1 Score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$$

$\geq 11$0

For quantization, the paper presents FP16 conversion conceptually as

$\geq 11$1

where $\geq 11$2 determines the precision range. Integer quantization is given by

$\geq 11$3

where $\geq 11$4 denotes tensor values, $\geq 11$5 the scale factor, and $\geq 11$6 the bit-width, such as 8 for INT8 and 4 for INT4. For INT4, the paper also gives an AWQ scaling update:

$\geq 11$7

where $\geq 11$8 is the step size and $\geq 11$9 is the quantization error or loss.

These definitions matter because the paper uses CP-AnemiC not only for full-precision classification but also for systematic comparison across FP32, FP16, INT8, and INT4 representations. The dataset is therefore used as a test bed for precision–performance trade-offs in mobile healthcare inference.

6. Reported performance, limitations, and deployment implications

The strongest validation result reported on CP-AnemiC occurred at epoch 26 in one fold: accuracy 0.9688, precision 0.9773, recall 0.9773, F1 score 0.9773, and AUC 0.9923 (Romero et al., 20 Jul 2025).

For final inference before quantization, the FP32 model achieved loss 0.2141, accuracy 0.9313, precision 0.9374, recall 0.9500, F1 score 0.9428, and AUC 0.9657. The corresponding post-training quantization results were reported as follows:

• FP16: loss 0.2149, accuracy 0.9250, precision 0.9370, recall 0.9400, F1 0.9377, AUC 0.9654.
• INT8: loss 0.7441, accuracy 0.7125, precision 0.7697, recall 0.7519, F1 0.7607, AUC 0.9005.
• INT4: loss 2.3136, accuracy 0.4313, precision 0.2000, recall 0.0100, F1 0.0196, AUC 0.6387.

Within the reported setup, FP16 preserves most of the FP32 model’s performance, whereas INT8 and especially INT4 show substantial degradation. The paper therefore supports the view that CP-AnemiC is suitable for evaluating lightweight deployment strategies, but also that aggressive quantization can materially harm diagnostic performance.

The paper identifies several limitations that are directly relevant to the dataset. First, the dataset size is still relatively small for deep learning by modern standards, which motivates the use of cross-validation and early stopping. Second, the class distribution is not perfectly balanced, and the gender distribution is also skewed. Third, the dataset is geographically limited to Ghana, which may constrain model generalization to other populations or clinical settings. Fourth, the paper does not report extensive bias analysis or detailed acquisition standardization, leaving site-wise and demographic confounding insufficiently characterized.

The deployment framing is explicit. The paper emphasizes mobile and edge-device suitability, uses MobileNet for efficiency, and studies post-training quantization to reduce memory footprint and improve inference efficiency. It also identifies future interest in devices such as NVIDIA Jetson Xavier NX and TX2 NX. In that context, CP-AnemiC functions as a clinically anchored pediatric benchmark for testing whether compact vision models can retain acceptable anemia-detection performance under constrained numerical precision.