Vision Mamba-Based Dual-phase Self-supervised Framework for Neonatal Jaundice Diagnosis

Neonatal jaundice is a common and potentially
serious condition that, if left undiagnosed or untreated, can
lead to severe neurological complications in newborns.
Existing diagnostic methods are often invasive and face
limitations in accuracy, accessibility, and data availability,
especially in resource-constrained environments. This study
introduces NeoViM, an adapted MambaVision-based
framework for neonatal jaundice classification. The
framework adopts a two-stage approach: In the first stage,
the adapted MambaVision is used as a deep feature extractor.
Self-Relative Clustering (SRC) is then applied to learn
discriminative features by organizing images into clinically
meaningful clusters, enabling effective unsupervised learning
from limited data. To further enhance feature quality, a selfsupervised learning strategy based on Linear Kernel
Centered Alignment (LCKA) loss is employed to refine the
extracted representations. In the second stage, the pretrained model is fine-tuned on a small labeled dataset,
allowing the system to adapt the learned features for accurate
jaundice classification. The methodology was evaluated using
the publicly available NJN dataset and achieved a
classification accuracy of 93.42% and an F1 score of 93.37%,
outperforming previous methods applied to the same dataset.
This two-step framework ensures high diagnostic
performance while maintaining scalability and accessibility
across diverse clinical settings.