Abstract: Anemia, defined by a reduction in red blood cell count or hemoglobin (Hb) concentration below normal levels, remains a significant global health issue, affecting approximately 30–40% of women and children worldwide. Early detection is critical to prevent complications such as fatigue, cognitive impairment in children, and increased maternal-fetal morbidity. Traditional diagnostic methods rely on invasive and resource-intensive blood tests, limiting accessibility in low-resource settings. This project proposes a novel, non-invasive approach for anemia detection using computer vision and deep learning techniques applied to smartphone-captured fingernail images. Visual signs such as nail-bed pallor and koilonychia (spoon-shaped nails), which correlate with hemoglobin deficiency, form the basis of this image-based screening system. The proposed framework includes assembling a labeled dataset of fingernail images paired with ground-truth Hb values from clinical and public sources. Preprocessing steps involve region of interest (ROI) extraction using YOLOv8 for fingernail detection, color normalization to account for skin tone variations and lighting inconsistencies, and data augmentation for robustness. A convolutional neural network (CNN) architecture, such as DenseNet169 or MobileNetV3, is fine-tuned for classification of anemic versus non-anemic cases. Explainable AI methods like Grad-CAM are employed to ensure model transparency and highlight relevant image regions influencing predictions. Deployment considerations include optimizing models for on-device inference using TensorFlow Lite, integrating real-time user guidance for image capture, and ensuring compliance with privacy and regulatory standards. This solution aims to democratize anemia screening, enabling scalable, accessible, and non-invasive early diagnosis in community and telehealth settings.

Keywords: Anemia detection, non-invasive, nail imaging, deep learning, computer vision, YOLOv8, DenseNet169, MobileNetV3, Explainable AI, Grad-CAM, telehealth.

Downloads: | DOI: 10.17148/IJIREEICE.2026.14392