Advancing Anemia Detection With Deep Neural Networks: A Comparative Analysis of Training Strategies Using Conjunctival Images

Anemia, characterized by decreased hemoglobin concentration and reduced red blood cell count, affects more than one-third of the global population and contributes substantially to illness and death among young children and pregnant women [1]. Conventional laboratory-based screening is resource-intensive and often inaccessible in low-resource settings [2]. Clinical assessment of conjunctival pallor offers a rapid, noninvasive alternative; however, its accuracy varies with observer experience and lighting conditions [3]. Deep convolutional neural networks can learn useful image features directly from data, improving the detection of subtle signs. Initial image-processing and thresholding methods achieved about 80%–84% accuracy in detecting conjunctival pallor [2, 3]. More recent studies using ensemble learning and CNNs on augmented data reported nearly 90% accuracy and added hemoglobin level prediction in pediatric groups [4-6]. However, previous studies have varied in algorithmic approach and dataset scope: Tamir et al. applied SVM thresholding to the anterior conjunctiva with 78.9% accuracy, Noor et al. compared decision trees, SVM, and KNN reaching up to 82.6% accuracy, Dimauro et al. mapped noninvasive devices, Appiahene CP-AnemiC yielded 80.6% accuracy, Appiahene's smartphone application reached 92.5% accuracy, Asare et al. reported 98.5% accuracy on pediatric conjunctiva, and Sehar et al. evaluated ensembles and CNNs achieving 89.5% accuracy (see Table S1 for the complete list of studies) [2-8]. A systematic review of these studies was conducted in accordance with PRISMA guidelines (see Figure S1). The question of whether training on comprehensive full-eye images rather than isolated palpebral views yields better generalization has not been directly addressed. We used 218 conjunctival photographs (126 healthy and 92 anemic) from the Eyes-Defy-Anemia dataset, which provides both full-eye photographs and corresponding isolated palpebral images [9]. Images were divided into training (n = 174; 80%), validation (n = 22; 10%), and test (n = 22; 10%) sets by stratified sampling (Figure S2). Labels were based on WHO hemoglobin thresholds (< 12 g/dL for nonpregnant females and < 13 g/dL for males) and incorporated biological sex. Gender, age, and Hb levels were provided, and the anemic labels followed WHO definitions [10-12]. Images were split using stratified sampling into training (n = 174), validation (n = 22), and test (n = 22) sets, with a fixed random seed (Figure S2). This mitigated the impact of random split variation and allowed consistent evaluation across models. Four AutoML image classification models were trained using Google Vertex AI (single-label, default settings), which performs transfer learning on a modern CNN backbone with automated hyperparameter tuning. We controlled data ingestion, labeling, augmentation, and evaluation, but the exact architecture is abstracted in AutoML. Four single-label classifiers were built with an AutoML platform. Models 1 and 2 were trained on isolated palpebral images shown in Figures S3 and S4. Model 1 was tested on palpebral images, and Model 2 was tested on full-eye images. Models 3 and 4 were trained on full-eye photographs shown in Figures S5 and S6. Model 3 tested on palpebral views, and Model 4 on full-eye views. Data augmentation (cropping, rotation, brightness adjustment, and noise) was applied, and all models were initialized with ImageNet-pretrained weights [4, 13]. Hyperparameters were tuned automatically, and training took approximately 1.5 h per model. Test set performance was evaluated using the area under the precision-recall curve (AUC), sensitivity, specificity, and Cohen's κ. Differences in AUC were compared using two-sample Z-tests (α = 0.05). Model performances on the test set are displayed in Table 1. Model 1 achieved an AUC of 0.765, sensitivity of 44%, specificity of 67%, and κ of 0.11. Model 2 achieved an AUC of 0.834, sensitivity of 78%, specificity of 83%, and κ of 0.62. Model 3 achieved an AUC of 0.899, sensitivity of 100%, specificity of 67%, and κ of 0.28. Model 4 achieved an AUC of 0.865, sensitivity of 44%, specificity of 100%, and κ of 0.44. Model 3 versus Model 1: ΔAUC = 0.134, Z = 2.205, p = 0.0275, Δκ = 0.17. Model 2 versus Model 4: ΔAUC = −0.031, Z = 6.63, p < 0.0001, Δκ = 0.62. Clinicians' evaluation of conjunctival pallor functions as an anemia screening tool, yet their assessments depend on their experience level and the available lighting conditions [3]. The first deep learning research in ophthalmology focused on multi-retinal disease identification and diabetic retinopathy screening, which led experts to recognize CNNs as models that could match expert-level image classification performance, thus inspiring our investigation of best training approaches [4, 5]. Models trained on entire eye images detect wider anatomical details, which produce superior results when analyzing palpebral images. The AUC value reached 0.899% and 100% sensitivity on palpebral tests for Model 3 when it learned from localized vascular cues and surrounding ocular anatomy, despite no training exposure to cropped views. Model 1 received training and testing only with cropped palpebral images, yet demonstrated 44% sensitivity and a Cohen's κ of 0.11 while failing to learn beyond its limited view. Model 2 achieved a κ of 0.62 after training on palpebral images and testing on full-eye photos, which shows that cross-view evaluation benefits from additional context. Model 4 achieved perfect specificity (100%) but low sensitivity (44%), which we attribute to several factors. First, non-diagnostic context dominance may contribute: full-eye images include sclera, skin, lids, and cornea, which can lead the model to overfit to lighting, skin tone, or gaze direction. Second, signal dilution occurs when resizing compresses the palpebral region, weakening key chromatic cues. Third, greater intra-class variance from full-eye views due to pose and exposure differences adds noise. Finally, the default threshold selected by AutoML prioritized overall accuracy, favoring specificity at the cost of sensitivity. The results show that training models with a broad range of image perspectives leads to better diagnostic performance in specific tasks. Our model outputs a continuous probability score that tracks anemia severity. Since palpebral pallor intensifies as hemoglobin declines, sensitivity improves for moderate-to-severe anemia, especially under recall-optimized thresholds. Mild cases, where visual features overlap more with healthy eyes, are more likely to be missed, consistent with known challenges in clinical pallor grading. The absence of k-fold cross-validation is a limitation due to dataset size and AutoML constraints. However, we address this by using a stratified split with a fixed seed, ensuring test set isolation. Sensitivity analyses using alternative seeds preserved model ranking and performance trends (see Limitations). Future iterations will implement full cross-validation in a custom pipeline. The current user interface and display of our models' functionality when images are uploaded for a diagnostic prediction are demonstrated in Figure 1. Future research must create training datasets encompassing multiple anatomical views, including full-face, eyelid, and conjunctival regions, to establish models with transferable feature representations. Implementing curriculum learning methods for networks to learn complex context levels successively could enhance their detection capabilities for specialized tasks. The research should continue to evaluate how different proportions of full-eye to palpebral images during training affect model sensitivity thresholds and false-positive rates, which could enable classifier confidence adjustments for various clinical situations. The integration of metadata such as age, skin tone, and prior hemoglobin measurements would enable model adaptation for patient-specific risk profiles to develop a decision support system from the current binary screening tool. The mechanistic analysis of performance enhancement requires researchers to understand whether models use scleral contrast patterns, vascular branching architectures, or tear film reflections. Standardized full-spectrum imaging under infrared or polarized light could improve feature contrast and model accuracy. The collaboration between high-resolution imaging techniques and biomechanical modeling of eyelid tissues could discover new biomarkers that extend the possibilities of noninvasive anemia detection. Building diverse datasets remains essential: Multi-institutional and cross-continental image repositories must be created to evaluate model resistance against camera hardware variations, environmental lighting and patient demographic differences. Prospective cohort studies following hemoglobin changes over time will validate DNN predictions to develop dynamic anemia progression and treatment response monitoring capabilities. Our approach promotes generalized training foundations supporting a transition from dataset siloing towards comprehensive imaging strategies emphasizing diverse contexts. This approach shows potential for application beyond anemia screening to other point-of-care diagnostics because it allows models to detect lesions while maintaining broad scene understanding, thus making AI tools adaptable and reliable in complex clinical settings. Practical applications involve taking full-eye images through smartphones or ophthalmic cameras before performing automated segmentation, followed by AI-based analysis. Our dataset's small size and limited demographic representation are compensated for using full-eye and palpebral images to create a scalable solution. Standardized image acquisition, expanded multi-ethnic datasets, and development of multi-task networks for anemia detection and hemoglobin estimation will be essential. Combining adaptive augmentation methods with exploring other body sites, such as nail beds, will improve noninvasive screening methods. Training DNN models with diverse segmentation methods that use full-eye context information produces the best results for anemia detection (AUC 0.899, sensitivity 100%) when applied to palpebral regions. The proposed method enables a cost-effective and scalable screening method for limited resource areas. Future research should focus on multi-ethnic validation, smartphone deployment, and multi-task modeling to improve accessible noninvasive anemia diagnostics. We would like to sincerely thank the creators of the Eyes Defy Anemia dataset and the researchers involved in its development and public release. Their contributions have enabled this project and advanced the field of accessible, AI-powered anemia detection. The authors have nothing to report. This study did not involve the use of human subjects or identifiable human data and was therefore exempt from institutional review board (IRB) approval. The authors have nothing to report. The authors declare no conflicts of interest. The data used in this study are publicly available through the Eyes Defy Anemia dataset repository upon request from the dataset creators or the hosting institution. Data S1: ajh70143-sup-0001-Supinfo.docx. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.