Comparative Study of CNN Backbones for Hemoglobin Estimation from Fingernail Images

Judita Rudokaite, L.L. Sharon Ong, Joris van Eijkelenburg, Katja van den Hurk, Elisabeth Huis in 't Veld

Track: Type A (Regular Papers)

Keywords: Hemoglobin measurement, Machine learning, Fingernails

Abstract: Non-invasive hemoglobin (Hb) screening from smartphone images could reduce reliance on finger-prick tests in blood donation and outpatient clinics. While previous studies have shown feasibility on small datasets, few have systematically examined how methodological choices affect performance. We compared four frozen convolutional backbones—ConvNeXt-Tiny, ResNet-50, EfficientNetV2-S, and MobileNetV3-Large—using fingernail crops under a shared pipeline. Deep embeddings and color features were extracted and combined through early and late fusion, evaluated with 5-fold stratified cross-validation. The best-performing model, ConvNeXt-Tiny, achieved an out-of-fold MAE of 0.6 mmol/L. Robustness analyses showed stable performance under illumination changes (MAE shift < 0.02~mmol/L), minimal subgroup differences by sex or donation type (≤ 0.04~mmol/L), and well-calibrated uncertainty estimates (interval width 1.0–1.2~mmol/L, 0.9~mmol/L with normalization). These findings demonstrate that Hb-related information can be extracted from self-captured fingernail images using lightweight, reproducible pipelines. While current performance does not yet meet clinical standards, the results support the feasibility of image-based Hb estimation and motivate larger, multi-site studies for clinical validation.

Serve As Reviewer: ~Judita_Rudokaite1

Submission Number: 17