一种基于Unet改进的SAM模型实现的肿瘤标注方法
By introducing the Unet architecture and multi-scale edge enhancement technology into the SAM model, the problem of poor tumor annotation performance of SAM in medical images is solved, and accurate annotation of tumors with complex edges and variable shapes is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG FIRST MEDICAL UNIV & SHANDONG ACADEMY OF MEDICAL SCI
- Filing Date
- 2025-04-22
- Publication Date
- 2026-07-17
AI Technical Summary
The Segement Anything Model (SAM) struggles to effectively distinguish complex boundaries between tumor lesions and background organs in medical images, especially weak boundaries and complex edges, resulting in poor annotation performance.
The Unet architecture is improved in the SAM model by adding a cascaded upsampling decoding module and edge enhancement attention. The annotation accuracy is improved by multi-scale edge enhancement technology, and edge information is enhanced by using Laplacian pyramid and edge enhancement attention. The model is trained by combining cross-entropy loss and DICE loss.
It improves the edge quality and accuracy of tumor annotation, especially for tumors with complex edges and varied shapes, and achieves higher quality region of interest annotation.
Smart Images
Figure CN120388166B_ABST