掷骰子
计算机科学
分割
人工智能
计算机视觉
忠诚
模式识别(心理学)
安全性令牌
边界(拓扑)
地铁列车时刻表
融合
变化(天文学)
基础(拓扑)
过程(计算)
简单(哲学)
图像分割
集合(抽象数据类型)
匹配(统计)
选择(遗传算法)
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-12-05
卷期号:: 538-541
标识
DOI:10.1109/icftic68075.2025.11324748
摘要
Accurate CT/MRI organ segmentation supports diagnosis, planning, and routine follow-up. The original MedDINOv3 performs well, but it still misses detail along thin or low-contrast boundaries and shows variation across modalities. We present an enhanced MedDINOv3 that adds a cross-scale attention fusion module. It aggregates token features from ViT blocks 2, 5, 8, and 11 with learned, data-dependent weights. Early tokens help with edges; deeper tokens keep shapes coherent. We also tighten a three-stage domain-adaptive pretraining schedule on CT-3M (3.868 million axial slices) so high-resolution cues survive augmentation. On AMOS22, our model reaches 88.62% mean Dice over 15 organs. That is +1.24 points over the original MedDINOv3 and +3.81 over nnU-Net. On BTCV, the mean Dice is 80.15% (+1.36 / +6.85). For tumors, we obtain 71.93% on KiTS23 and 76.54% on LiTS, with steady gains over the base model. The design keeps the decoder light and adds only minor overhead. Boundary fidelity improves, global structure remains stable, and cross-modality behavior is more consistent. The approach is simple to train and practical for clinical workflows.
科研通智能强力驱动
Strongly Powered by AbleSci AI