Objectives Plain abdominal radiography is a widely used imaging modality for diagnosing neonatal necrotizing enterocolitis (NEC), but the characteristic features of stage I NEC are often subtle, making early diagnosis challenging. This study explores the application of deep learning (DL) models to assist in the early diagnosis of stage I NEC. Materials and methods This retrospective study included 380 and 300 neonates who underwent abdominal radiography at two centers between June 2016 and December 2023. Neonates were grouped based on a diagnosis of stage I NEC. DL features were extracted from the radiographs using the DenseNet121 model, based on which radiomics models were constructed using logistic regression (LR) and random forest (RF) algorithms. Performance was evaluated through receiver operating characteristic (ROC) curves. Both the training and external validation cohorts were used to assess model accuracy in distinguishing stage I NEC. Additionally, a direct comparison with human expert diagnostic performance was conducted. Results In the training cohort, 25 DL features were selected for model development. The area under the ROC curve (AUC) for LR and RF models was 0.972 (95% CI: 0.956–0.988) and 0.961 (95% CI: 0.942–0.980), respectively. In the external validation cohort, the models demonstrated AUCs of 0.964 (95% CI: 0.943–0.986) and 0.951 (95% CI: 0.925–0.976), respectively. These models evidently outperformed human experts in diagnostic performance. Conclusion The DL model based on plain abdominal radiography effectively identified stage I NEC in neonates. This approach offers a non-invasive method to enhance early NEC diagnosis and support clinical decision-making. Key Points <Emphasis Type="BoldItalic">Question</Emphasis>Deep learning (DL) models applied to plain abdominal radiography can enhance the early diagnosis of stage I neonatal necrotizing enterocolitis (NEC). <Emphasis Type="BoldItalic">Findings</Emphasis>In this retrospective study involving 680 neonates from two centers, DL-based radiomics models achieved much higher accuracy for diagnosing stage I NEC than human radiologists. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis>DL models based on plain abdominal radiography have the ability to significantly improve the early identification of stage I NEC, offering a non-invasive tool to support radiologists in early diagnosis. Graphical Abstract

Deep learning feature-based model on abdominal radiography outperforms experts for early necrotizing enterocolitis diagnosis in neonates

期刊论文
发表日期：

2026年1月13日

查看全文

Yu Wu,
Hao Yang,
Xiaomei Luo,
Haige Zheng,
Yan Zhou,
Chengyan Chen,
Rui Wang,
Hongsheng Liu,
Liandong Zuo,
Xiaochun Zhang,
Kejian Wang,
Xuehua Peng

摘要

Objectives

Plain abdominal radiography is a widely used imaging modality for diagnosing neonatal necrotizing enterocolitis (NEC), but the characteristic features of stage I NEC are often subtle, making early diagnosis challenging. This study explores the application of deep learning (DL) models to assist in the early diagnosis of stage I NEC.

Materials and methods

This retrospective study included 380 and 300 neonates who underwent abdominal radiography at two centers between June 2016 and December 2023. Neonates were grouped based on a diagnosis of stage I NEC. DL features were extracted from the radiographs using the DenseNet121 model, based on which radiomics models were constructed using logistic regression (LR) and random forest (RF) algorithms. Performance was evaluated through receiver operating characteristic (ROC) curves. Both the training and external validation cohorts were used to assess model accuracy in distinguishing stage I NEC. Additionally, a direct comparison with human expert diagnostic performance was conducted.

Results

In the training cohort, 25 DL features were selected for model development. The area under the ROC curve (AUC) for LR and RF models was 0.972 (95% CI: 0.956–0.988) and 0.961 (95% CI: 0.942–0.980), respectively. In the external validation cohort, the models demonstrated AUCs of 0.964 (95% CI: 0.943–0.986) and 0.951 (95% CI: 0.925–0.976), respectively. These models evidently outperformed human experts in diagnostic performance.

Conclusion

The DL model based on plain abdominal radiography effectively identified stage I NEC in neonates. This approach offers a non-invasive method to enhance early NEC diagnosis and support clinical decision-making.

Key Points

QuestionDeep learning (DL) models applied to plain abdominal radiography can enhance the early diagnosis of stage I neonatal necrotizing enterocolitis (NEC).

FindingsIn this retrospective study involving 680 neonates from two centers, DL-based radiomics models achieved much higher accuracy for diagnosing stage I NEC than human radiologists.

Clinical relevanceDL models based on plain abdominal radiography have the ability to significantly improve the early identification of stage I NEC, offering a non-invasive tool to support radiologists in early diagnosis.

Graphical Abstract