一种基于可学习令牌的医学图像分割多模型聚合方法
By using a learnable token-based approach, combined with pseudo-label synthesis and feature layer and output layer distillation loss optimization, the problems of poor model scalability and poor segmentation performance in multi-task learning are solved, achieving more efficient multi-model knowledge fusion and accurate medical image segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI ARTIFICIAL INTELLIGENCE INNOVATION CENT
- Filing Date
- 2024-09-27
- Publication Date
- 2026-07-17
AI Technical Summary
Existing multi-task learning methods suffer from poor scalability, catastrophic forgetting, and poor segmentation performance in medical image multi-organ segmentation model aggregation tasks. In particular, semi-supervised methods that rely solely on pseudo-labels lead to class imbalance, and knowledge distillation methods that depend on the output layer cannot fully transfer knowledge.
A learnable token-based approach is adopted to synthesize pseudo-labels from multiple teacher models and combine them with real new class labels to form a fully supervised label map for the student model. Learnable tokens are then used to transform the feature distribution of the student model. The student model parameters are optimized by combining distillation partial loss, feature distillation/alignment partial loss, and segmentation partial loss, thereby achieving knowledge fusion between the feature layer and the output layer.
This enables the simultaneous processing of multiple tasks without relying on old training data, improving the accuracy and scalability of medical image segmentation models, enhancing the effect of multi-model knowledge fusion, and particularly improving segmentation performance on cross-domain datasets.
Smart Images
Figure CN119206228B_ABST