Learning uncertainty fields for joint brain stroke diagnosis and intrinsic lesion localization in CT and MRI

Early stroke changes can be difficult to detect and localize on neuroimaging, particularly when lesion boundaries are indistinct. We developed the Dynamic Fuzzy Encoder Network (DFE-Net), which uses learnable fuzzy membership fields to represent spatial uncertainty and jointly classify stroke and localize lesions. We evaluated the model on a multicentre computed tomography (CT) cohort comprising 6650 patients and 8873 images and a separate magnetic resonance imaging (MRI) cohort comprising 67 patients and 615 images. Training and testing were performed separately within each modality. DFE-Net achieved an area under the receiver operating characteristic curve of 0.974 and an F1 score of 0.897 on CT, and corresponding values of 0.978 and 0.879 on MRI. In the CT cohort, lesion localization against expert annotations yielded a Dice score of 0.70. MRI localization was assessed using gradient-derived pseudo-targets because expert lesion masks were unavailable. Expected calibration error was ~0.022 in both cohorts, and the model contained 0.20 million parameters. These results support joint diagnosis and spatial localization with a compact model, while prospective validation and expert-annotated MRI evaluation remain necessary.

Authors

Institutions

Publication Details

Journal
npj Digital Medicine
Published
2026-11-14
DOI
https://doi.org/10.1038/s41746-026-03458-9
Primary Topic
Medical Image Segmentation Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Learning uncertainty fields for joint brain stroke diagnosis and intrinsic lesion localization in CT and MRI

Xuanfeng Qin, Hao Sun, Kangwei Wang, Xiaojun Dong et al.
1 links
npj Digital Medicine
Medical Image Segmentation Techniques
article

Learning uncertainty fields for joint brain stroke diagnosis and intrinsic lesion localization in CT and MRI

Xuanfeng Qin, Hao Sun, Kangwei Wang, Xiaojun Dong, Zehao Jin, Jieyi Yang, Shuzhen Xu
article en

Abstract

Early stroke changes can be difficult to detect and localize on neuroimaging, particularly when lesion boundaries are indistinct. We developed the Dynamic Fuzzy Encoder Network (DFE-Net), which uses learnable fuzzy membership fields to represent spatial uncertainty and jointly classify stroke and localize lesions. We evaluated the model on a multicentre computed tomography (CT) cohort comprising 6650 patients and 8873 images and a separate magnetic resonance imaging (MRI) cohort comprising 67 patients and 615 images. Training and testing were performed separately within each modality. DFE-Net achieved an area under the receiver operating characteristic curve of 0.974 and an F1 score of 0.897 on CT, and corresponding values of 0.978 and 0.879 on MRI. In the CT cohort, lesion localization against expert annotations yielded a Dice score of 0.70. MRI localization was assessed using gradient-derived pseudo-targets because expert lesion masks were unavailable. Expected calibration error was ~0.022 in both cohorts, and the model contained 0.20 million parameters. These results support joint diagnosis and spatial localization with a compact model, while prospective validation and expert-annotated MRI evaluation remain necessary.

npj Digital Medicine
Ningbo University (CN), Fujian University of Traditional Chinese Medicine (CN), Huaihua University (CN), Wenzhou Medical University (CN), Hua Medicine (China) (CN), Huashan Hospital (CN), Yueqing People's Hospital (CN), University College London (GB), Technical University of Munich (DE)
Good health and well-being
Medical Image Segmentation Techniques
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