Machine vision model using nail images for non-invasive detection of iron deficiency anemia in university students

• Published in: Frontiers in big data Vol. 8; p. 1557600
• Main Authors: Navarro-Cabrera, Jorge Raul, Valles-Coral, Miguel Angel, Farro-Roque, María Elena, Reátegui-Lozano, Nelly, Arévalo-Fasanando, Lolita
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 09.04.2025
• Subjects: Big Data; cross validation; data capture; deep learning; DenseNet169; non-invasive diagnostics; deep learning; DenseNet169; non-invasive diagnostics; cross validation; data capture
• ISSN: 2624-909X; 2624-909X
• DOI: 10.3389/fdata.2025.1557600

Iron deficiency anemia (IDA) is a global health issue that significantly affects quality of life. Non-invasive methods, such as image analysis using artificial vision, offer accessible alternatives for diagnosis. This study proposes a DenseNet169-based model to detect anemia from nail images and compares its performance with that of the Rad-67 hemoglobin meter. A cross-sectional study was conducted with 909 nail images collected from university students aged 18-25 years at the Universidad Nacional de San Martín, Peru. Samsung Galaxy A73 5G was used to capture images under controlled conditions, and clinical data were complemented with hemoglobin readings from the Rad-67 device. The images were pre-processed using segmentation and data augmentation techniques to standardize the dataset. Three models (DenseNet169, InceptionV3, and Xception) were trained and evaluated using metrics, such as accuracy, recall, and AUC. DenseNet169169 demonstrated the best performance, achieving an accuracy of 0.6983, recall of 0.6477, F1-Score of 0.6525, and AUC of 0.7409. Despite the presence of false-negatives, the results showed a positive correlation with Rad-67 readings. The DenseNet169-based model proved to be a promising tool for non-invasive detection of iron deficiency anemia, with potential for application in clinical and educational settings. Future improvements in preprocessing and dataset diversification could enhance performance and applicability.