Deep learning grading of Schistosoma japonicum-associated liver fibrosis from ultrasound images: a weighted soft-voting ensemble of convolutional and transformer networks
- 1Chinese Center for Disease Control and Prevention (Chinese Center for Tropical Diseases Research); School of Global Health, Chinese Center for Tropical Diseases
Abstract
Background: Schistosomiasis morbidity still persists in China despite extensive progress towards elimination of new and active infections. Diagnosis of liver fibrosis related to Schistosoma japonicum remains challenging due to the subjective nature of ultrasound grading and the shortage of experienced sonographers in endemic counties. Methods: We developed a deep learning model that grades schistosomal liver fibrosis from a single B-mode ultrasound image without manual delineation of a region of interest. We used 2,921 images from 2,921 adults (aged 18-88 years, median age 54 years; 46.4% female) from a schistosomiasis hospital in Jiangxi Province, China, between 2019 and 2023. Of these, 12.9%, 34.6%, 32.4% and 20.2% had grade 0, 1, 2 and 3 fibrosis, respectively. Fibrosis grades were based on the clinical ultrasound record and re-evaluation by an experienced radiologist according to the Chinese Handbook of Schistosomiasis Control; discordant cases were reviewed by a second radiologist and resolved by consensus. Predictions from four neural networks with different architectures (ResNet-18, EfficientNet-B0, DenseNet-121 and Swin-T) were combined by weighted soft voting, with weights derived from validation performance. Results: For the four-grade task this model achieved a macro-F1 of 0.8527 (95% CI 0.8184-0.8856), an accuracy of 0.8500 (0.8136-0.8818) and a quadratic weighted kappa of 0.9187 (0.8983-0.9373) on a fixed test set of 440 images. For detecting significant fibrosis (grade 2 or 3; test-set prevalence 0.530), sensitivity was 0.9700, specificity was 0.9179, positive predictive value was 0.9300, negative predictive value was 0.9645, and AUROC was 0.9817. The ensemble achieved a higher macro-F1 than each individual backbone. Every ensemble misclassification was to an adjacent grade; no image with severe fibrosis was classified as no or mild fibrosis, or vice versa, whereas each individual backbone made 1-4 errors spanning two or more grades. Conclusions: The model may provide an automated aid for Schistosoma japonicum-associated fibrosis grading and could potentially support screening and follow-up in endemic areas after external validation.
Related
- Effects of Helleborus thibetanus extract, a novel plant-derived molluscicide on Oncomelania hupensis, the intermediate host of Schistosoma japonicum
- MedFuse: dual-stream fusion of convolutional and vision transformer-based features for enhanced medical image classification
- Interpretable attention-deficit hyperactivity disorder diagnosis from functional magnetic resonance imaging via hybrid convolutional neural network–transformer learning and multi-level graph-based comprehensibility
- Comparative evaluation of convolutional and transformer-based object detection models for brain tumor localization in MRI
- Evaluation of inter-observer agreement for periportal fibrosis related to Schistosoma mansoni: A population-based study of 3302 individuals using point-of-care ultrasound images and video in the SchistoTrack cohort
BibTeX
@misc{guo2026deep,
title = {{Deep learning grading of Schistosoma japonicum-associated liver fibrosis from ultrasound images: a weighted soft-voting ensemble of convolutional and transformer networks}},
author = {Z. Guo and E. S. Ockenden and M. Wei and Q. Li and S. Li and J. Xue and Y. Li and D. Lin and X. Zhou and G. F. Chami and S. Li},
howpublished = {medRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.02.26364999},
url = {https://www.medrxiv.org/content/10.64898/2026.11.02.26364999v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Deep learning grading of Schistosoma japonicum-associated liver fibrosis from ultrasound images: a weighted soft-voting ensemble of convolutional and transformer networks
AU - Z. Guo
AU - E. S. Ockenden
AU - M. Wei
AU - Q. Li
AU - S. Li
AU - J. Xue
AU - Y. Li
AU - D. Lin
AU - X. Zhou
AU - G. F. Chami
AU - S. Li
PY - 2026
DA - 2026/11/02
PB - medRxiv
DO - 10.64898/2026.11.02.26364999
UR - https://www.medrxiv.org/content/10.64898/2026.11.02.26364999v1
AB - Background: Schistosomiasis morbidity still persists in China despite extensive progress towards elimination of new and active infections. Diagnosis of liver fibrosis related to Schistosoma japonicum remains challenging due to the subjective nature of ultrasound grading and the shortage of experienced sonographers in endemic counties. Methods: We developed a deep learning model that grades schistosomal liver fibrosis from a single B-mode ultrasound image without manual delineation of a region of interest. We used 2,921 images from 2,921 adults (aged 18-88 years, median age 54 years; 46.4% female) from a schistosomiasis hospital in Jiangxi Province, China, between 2019 and 2023. Of these, 12.9%, 34.6%, 32.4% and 20.2% had grade 0, 1, 2 and 3 fibrosis, respectively. Fibrosis grades were based on the clinical ultrasound record and re-evaluation by an experienced radiologist according to the Chinese Handbook of Schistosomiasis Control; discordant cases were reviewed by a second radiologist and resolved by consensus. Predictions from four neural networks with different architectures (ResNet-18, EfficientNet-B0, DenseNet-121 and Swin-T) were combined by weighted soft voting, with weights derived from validation performance. Results: For the four-grade task this model achieved a macro-F1 of 0.8527 (95% CI 0.8184-0.8856), an accuracy of 0.8500 (0.8136-0.8818) and a quadratic weighted kappa of 0.9187 (0.8983-0.9373) on a fixed test set of 440 images. For detecting significant fibrosis (grade 2 or 3; test-set prevalence 0.530), sensitivity was 0.9700, specificity was 0.9179, positive predictive value was 0.9300, negative predictive value was 0.9645, and AUROC was 0.9817. The ensemble achieved a higher macro-F1 than each individual backbone. Every ensemble misclassification was to an adjacent grade; no image with severe fibrosis was classified as no or mild fibrosis, or vice versa, whereas each individual backbone made 1-4 errors spanning two or more grades. Conclusions: The model may provide an automated aid for Schistosoma japonicum-associated fibrosis grading and could potentially support screening and follow-up in endemic areas after external validation.
ER -