TY - GEN
T1 - ISSD-NET
T2 - 32nd International Conference on Neural Information Processing, ICONIP 2025
AU - Li, Haowen
AU - Huang, Guoheng
AU - Yuan, Xiaochen
AU - Li, Yan
AU - Pun, Chi Man
AU - Lei, Baiying
AU - Huang, Zhixin
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.
PY - 2026
Y1 - 2026
N2 - Medical image segmentation heavily depends on labeled data, but its acquisition is costly and time-consuming. Semi-supervised learning (SSL) utilizes unlabeled data to relieve this burden, yet existing methods suffer from inaccurate pseudo-labels caused by distribution mismatch, hard label information loss, and inter-class size imbalance. To address these issues, we propose an advanced semi-supervised framework for medical image segmentation (ISSD-NET). ISSD-NET incorporates three key components: First, the Grad-CAM Splicing Module (GCSM) leverages Grad-CAM to locate semantically rich regions and splice them into augmented samples, aligning labeled and unlabeled data distributions for more reliable pseudo-labels. Second, the Self-Distillation Consistency Regularization Module (SCRM) generates soft pseudo-labels as supplementary supervision, preserving uncertainty and fine-grained details to overcome hard label limitations. Additionally, the Adaptive q-vMF Dice Loss dynamically adjusts similarity metrics to address class imbalance, further refining pseudo-label quality. Experiments on ACDC and LA datasets show that ISSD-NET outperforms state-of-the-art methods in semi-supervised medical segmentation, enhancing accuracy and offering a robust solution for reliable medical image analysis.
AB - Medical image segmentation heavily depends on labeled data, but its acquisition is costly and time-consuming. Semi-supervised learning (SSL) utilizes unlabeled data to relieve this burden, yet existing methods suffer from inaccurate pseudo-labels caused by distribution mismatch, hard label information loss, and inter-class size imbalance. To address these issues, we propose an advanced semi-supervised framework for medical image segmentation (ISSD-NET). ISSD-NET incorporates three key components: First, the Grad-CAM Splicing Module (GCSM) leverages Grad-CAM to locate semantically rich regions and splice them into augmented samples, aligning labeled and unlabeled data distributions for more reliable pseudo-labels. Second, the Self-Distillation Consistency Regularization Module (SCRM) generates soft pseudo-labels as supplementary supervision, preserving uncertainty and fine-grained details to overcome hard label limitations. Additionally, the Adaptive q-vMF Dice Loss dynamically adjusts similarity metrics to address class imbalance, further refining pseudo-label quality. Experiments on ACDC and LA datasets show that ISSD-NET outperforms state-of-the-art methods in semi-supervised medical segmentation, enhancing accuracy and offering a robust solution for reliable medical image analysis.
KW - dice loss
KW - medical image segmentation
KW - self knowledge distillation
KW - Semi-supervised learning
UR - https://www.scopus.com/pages/publications/105022786460
U2 - 10.1007/978-981-95-4378-6_34
DO - 10.1007/978-981-95-4378-6_34
M3 - Conference contribution
AN - SCOPUS:105022786460
SN - 9789819543779
T3 - Lecture Notes in Computer Science
SP - 486
EP - 500
BT - Neural Information Processing - 32nd International Conference, ICONIP 2025, Proceedings
A2 - Taniguchi, Tadahiro
A2 - Leung, Chi Sing Andrew
A2 - Kozuno, Tadashi
A2 - Yoshimoto, Junichiro
A2 - Mahmud, Mufti
A2 - Doborjeh, Maryam
A2 - Doya, Kenji
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 20 November 2025 through 24 November 2025
ER -