Image-Based Bilirubin Estimation for Non-Invasive Neonatal Jaundice Assessment Using Swin Transformer V2

Abstract Neonatal jaundice, is a prevalent and generally a benign condition in neonates that involves elevated levels of bilirubin, resulting in the yellowish discoloration of the skin and eyes. Accurate and timely measurement of bilirubin concentration is required to diagnose jaundiced neonates and to prevent serious complications resulting from untreated hyperbilirubinemia. Conventional bilirubin estimation methods commonly involve invasive blood sampling, indicating the need for reliable, accessible, and non-invasive image-based methods for bilirubin assessment. Therefore, there is a pressing need for a non-invasive, rapid, accessible, and explainable method for early bilirubin estimation and jaundice assessment. This study presents a Non-Invasive Bilirubin Estimation and Assessment Model (N-BEAM) for neonatal jaundice using Swin Transformer V2 with an explainable deep learning strategy. The presented N- BEAM model aims to estimate continuous Total Serum Bilirubin (TSB) values and provide a four-level assessment of neonatal jaundice without requiring blood samples during inference. Swin Transformer V2 is employed to extract hierarchical deep features from the preprocessed neonatal skin images through window-based self-attention and shifted-window mechanisms. Besides, global average pooling and a fully connected regression network is employed to predict the continuous TSB value in mg/dL. Subsequently, the predicted TSB value is converted into four study-defined categories, namely Normal, Mild, Moderate, and Severe, providing an interpretable bilirubin assessment without using a separate classification model. Moreover, Grad-CAM++ is integrated into the proposed framework to provide visual interpretability by highlighting the neonatal skin regions that contribute to the predicted TSB value. To validate the effectiveness of the proposed N-BEAM model, extensive experiments are conducted using two independent neonatal jaundice image datasets, including the NJN dataset and the NeoJaundice dataset, under a cross-dataset evaluation protocol in which one dataset is used for training and the other as an independent testing dataset. The experimental results infer the capability of the N-BEAM framework to provide non-invasive bilirubin estimation, severity assessment, and visually interpretable predictions from neonatal images.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-10963208/v1
Primary Topic
Neonatal Health and Biochemistry
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Image-Based Bilirubin Estimation for Non-Invasive Neonatal Jaundice Assessment Using Swin Transformer V2

Hanan Muhajab, Areej Muhajab
3 links
Research Square
Neonatal Health and Biochemistry
preprint

Image-Based Bilirubin Estimation for Non-Invasive Neonatal Jaundice Assessment Using Swin Transformer V2

Hanan Muhajab, Areej Muhajab
preprint en

Abstract

Abstract Neonatal jaundice, is a prevalent and generally a benign condition in neonates that involves elevated levels of bilirubin, resulting in the yellowish discoloration of the skin and eyes. Accurate and timely measurement of bilirubin concentration is required to diagnose jaundiced neonates and to prevent serious complications resulting from untreated hyperbilirubinemia. Conventional bilirubin estimation methods commonly involve invasive blood sampling, indicating the need for reliable, accessible, and non-invasive image-based methods for bilirubin assessment. Therefore, there is a pressing need for a non-invasive, rapid, accessible, and explainable method for early bilirubin estimation and jaundice assessment. This study presents a Non-Invasive Bilirubin Estimation and Assessment Model (N-BEAM) for neonatal jaundice using Swin Transformer V2 with an explainable deep learning strategy. The presented N- BEAM model aims to estimate continuous Total Serum Bilirubin (TSB) values and provide a four-level assessment of neonatal jaundice without requiring blood samples during inference. Swin Transformer V2 is employed to extract hierarchical deep features from the preprocessed neonatal skin images through window-based self-attention and shifted-window mechanisms. Besides, global average pooling and a fully connected regression network is employed to predict the continuous TSB value in mg/dL. Subsequently, the predicted TSB value is converted into four study-defined categories, namely Normal, Mild, Moderate, and Severe, providing an interpretable bilirubin assessment without using a separate classification model. Moreover, Grad-CAM++ is integrated into the proposed framework to provide visual interpretability by highlighting the neonatal skin regions that contribute to the predicted TSB value. To validate the effectiveness of the proposed N-BEAM model, extensive experiments are conducted using two independent neonatal jaundice image datasets, including the NJN dataset and the NeoJaundice dataset, under a cross-dataset evaluation protocol in which one dataset is used for training and the other as an independent testing dataset. The experimental results infer the capability of the N-BEAM framework to provide non-invasive bilirubin estimation, severity assessment, and visually interpretable predictions from neonatal images.

Research Square
Taif University (SA), Jazan University (SA)
Good health and well-being
Neonatal Health and Biochemistry
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon