Scholarlink

Background <p>Given the lack of research on deep learning for evaluating liver fibrosis in biliary atresia (BA), and the challenge of balancing accuracy and efficiency in lightweight models, this study aims to develop a liver fibrosis assessment model utilizing Dynamic SpatialLite-EfficientNet (DSL-Net), based on the BA-specific staging system.</p> Methods <p>A total of 530 HE whole slide images scanned from liver biopsies at Kasai portoenterostomy were used for the train (80%), validation (10%), and test (10%) of DSL-Net model. In the test cohort, the quantitative performance of DSL-Net was compared with that of existing lightweight models. Additionally, the model’s performance in liver fibrosis assessment was compared with that of six pathologists, and its prognostic predictive ability was further validated against traditional liver fibrosis scoring systems.</p> Results <p>The quantitative performance of DSL-Net model was 6.9&#xa0;M (Params), 0.5G (FLOPs), and 86.4 ± 0.7% (average Dice coefficient). It achieved an AUC of 0.906 for liver fibrosis assessment (≥ stage 3), comparable to two senior pediatric pathologists. With DSL-Net’s assistance, the AUCs for two junior pathologists and one attending pediatric pathologist improved from 0.685, 0.867, and 0.837 to 0.916 (<i>P</i> = 0.008), 0.907 (<i>P</i> = 0.005), and 0.948 (<i>P</i> = 0.021), respectively. Compared to the Metavir and Ishark scoring systems, DSL-Net model also demonstrated promising potential in predicting native liver survival (NLS), with 2-year and 5-year NLS AUCs of 0.8232 (<i>P</i> = 0.013; <i>P</i> = 0.046) and 0.7845 (<i>P</i> = 0.144; <i>P</i> = 0.656).</p> Conclusion <p>Based on the results of this study, the DSL-Net model demonstrates relatively higher accuracy in liver fibrosis assessment (≥ stage 3) and short-term prognostic prediction of BA, with both parameter count and computational cost comparable to the mainstream lightweight architecture. This positions DSL-Net as a promising, lightweight, and efficient tool for liver fibrosis assessment in BA, thereby better guiding clinical decision-making for BA patients.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development of a lightweight deep learning model for accurate assessment of liver fibrosis in biliary atresia

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Qianhui Yang,
  • Yuqiang Chen,
  • Yu Meng,
  • Xizhe Luo,
  • Li Zhao,
  • Xiaodan Xu,
  • Jichang Guo,
  • Shufang Zhang,
  • Jianghua Zhan

摘要

Background

Given the lack of research on deep learning for evaluating liver fibrosis in biliary atresia (BA), and the challenge of balancing accuracy and efficiency in lightweight models, this study aims to develop a liver fibrosis assessment model utilizing Dynamic SpatialLite-EfficientNet (DSL-Net), based on the BA-specific staging system.

Methods

A total of 530 HE whole slide images scanned from liver biopsies at Kasai portoenterostomy were used for the train (80%), validation (10%), and test (10%) of DSL-Net model. In the test cohort, the quantitative performance of DSL-Net was compared with that of existing lightweight models. Additionally, the model’s performance in liver fibrosis assessment was compared with that of six pathologists, and its prognostic predictive ability was further validated against traditional liver fibrosis scoring systems.

Results

The quantitative performance of DSL-Net model was 6.9 M (Params), 0.5G (FLOPs), and 86.4 ± 0.7% (average Dice coefficient). It achieved an AUC of 0.906 for liver fibrosis assessment (≥ stage 3), comparable to two senior pediatric pathologists. With DSL-Net’s assistance, the AUCs for two junior pathologists and one attending pediatric pathologist improved from 0.685, 0.867, and 0.837 to 0.916 (P = 0.008), 0.907 (P = 0.005), and 0.948 (P = 0.021), respectively. Compared to the Metavir and Ishark scoring systems, DSL-Net model also demonstrated promising potential in predicting native liver survival (NLS), with 2-year and 5-year NLS AUCs of 0.8232 (P = 0.013; P = 0.046) and 0.7845 (P = 0.144; P = 0.656).

Conclusion

Based on the results of this study, the DSL-Net model demonstrates relatively higher accuracy in liver fibrosis assessment (≥ stage 3) and short-term prognostic prediction of BA, with both parameter count and computational cost comparable to the mainstream lightweight architecture. This positions DSL-Net as a promising, lightweight, and efficient tool for liver fibrosis assessment in BA, thereby better guiding clinical decision-making for BA patients.