TY - JOUR
T1 - UiL-Net
T2 - Uncertainty-in-the-Loop Dual-Domain Learning for Reliable Skin Lesion Segmentation
AU - Dai, Duwei
AU - Dong, Caixia
AU - Dai, Guowei
AU - Huang, Haolin
AU - Yang, Xu
AU - Ren, Pengyu
AU - Wang, Liang
AU - Liu, Fan
AU - Yan, Qingsen
AU - Huang, Lanxuan
AU - Zeng, Wei
N1 - Publisher Copyright:
© 1991-2012 IEEE.
PY - 2026
Y1 - 2026
N2 - Reliable skin lesion segmentation is the cornerstone of dermoscopic Computer-Aided Diagnosis (CAD). Although spatial-frequency dual-domain architectures have enhanced segmentation performance by capturing complementary information, existing mainstream networks suffer from severe statistical miscalibration, where predicted confidence significantly deviates from the posterior probability distribution. This leads to models being blindly confident in uncertain predictions, hindering the provision of accurate segmentation results and trustworthy clinical risk pre-warning. To address this, we propose the uncertainty-in-the-loop dual-domain network, termed UiL-Net. Mirroring the clinical diagnostic workflow, this method constructs a two-pass “glance-and-gaze” mechanism: it first quantifies initial uncertainty through a global glance, and subsequently utilizes this prior to drive a gaze phase, which transforms uncertainty from a passive evaluation metric into an active guidance signal embedded throughout the entire dual-domain learning lifecycle. Within this closed-loop architecture, we design a trinity of uncertainty-guided components, comprising an Uncertainty-Guided Bidirectional Feature Fusion (UGBFF) module, an Uncertainty-Guided Graph Refinement (UGGR) module, and an Uncertainty-Guided Margin-Adaptive Loss (UGML) function. This synergistic suite operates across feature fusion, topological reasoning, and adaptive optimization to facilitate the network’s self-diagnosis and self-rectification. Extensive experiments on four public benchmarks (ISIC2017, ISIC2018, PH2, and HAM10000) demonstrate that UiL-Net not only achieves superior segmentation performance and statistical calibration but also produces uncertainty distributions that effectively reflect clinical diagnostic ambiguity, thereby enabling reliable risk pre-warning.
AB - Reliable skin lesion segmentation is the cornerstone of dermoscopic Computer-Aided Diagnosis (CAD). Although spatial-frequency dual-domain architectures have enhanced segmentation performance by capturing complementary information, existing mainstream networks suffer from severe statistical miscalibration, where predicted confidence significantly deviates from the posterior probability distribution. This leads to models being blindly confident in uncertain predictions, hindering the provision of accurate segmentation results and trustworthy clinical risk pre-warning. To address this, we propose the uncertainty-in-the-loop dual-domain network, termed UiL-Net. Mirroring the clinical diagnostic workflow, this method constructs a two-pass “glance-and-gaze” mechanism: it first quantifies initial uncertainty through a global glance, and subsequently utilizes this prior to drive a gaze phase, which transforms uncertainty from a passive evaluation metric into an active guidance signal embedded throughout the entire dual-domain learning lifecycle. Within this closed-loop architecture, we design a trinity of uncertainty-guided components, comprising an Uncertainty-Guided Bidirectional Feature Fusion (UGBFF) module, an Uncertainty-Guided Graph Refinement (UGGR) module, and an Uncertainty-Guided Margin-Adaptive Loss (UGML) function. This synergistic suite operates across feature fusion, topological reasoning, and adaptive optimization to facilitate the network’s self-diagnosis and self-rectification. Extensive experiments on four public benchmarks (ISIC2017, ISIC2018, PH2, and HAM10000) demonstrate that UiL-Net not only achieves superior segmentation performance and statistical calibration but also produces uncertainty distributions that effectively reflect clinical diagnostic ambiguity, thereby enabling reliable risk pre-warning.
KW - Dual-domain learning
KW - Reliable prediction
KW - Skin lesion segmentation
KW - Uncertainty guidance mechanism
UR - https://www.scopus.com/pages/publications/105053777643
U2 - 10.1109/TCSVT.2026.3741805
DO - 10.1109/TCSVT.2026.3741805
M3 - 文章
AN - SCOPUS:105053777643
SN - 1051-8215
JO - IEEE Transactions on Circuits and Systems for Video Technology
JF - IEEE Transactions on Circuits and Systems for Video Technology
ER -