Skin lesion image segmentation system and method based on multi-scale information reservation mechanism

By introducing distance decay constraints and feature fusion modules into multi-head attention, the problem of insufficient multi-scale information coordination in deep learning models for skin lesion image segmentation is solved, achieving more accurate and stable lesion boundary segmentation.

CN122289677APending Publication Date: 2026-06-26SHANDONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
SHANDONG UNIV
Filing Date
2026-03-10
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing deep learning models based on encoder-decoder structures struggle to effectively coordinate the relationship between multi-scale semantic information and spatial detail information in skin lesion image segmentation, resulting in inaccurate segmentation boundaries, blurred results, and unstable segmentation performance.

Method used

By introducing distance-based attenuation constraints into multi-head attention and setting attention heads with differentiated attenuation coefficients, combined with a feature fusion module, the preservation and transmission of multi-scale information are enhanced. In particular, the feature fusion module is set at the bottleneck position between the encoder and decoder to aggregate multi-scale contextual information.

Benefits of technology

It improves the ability to delineate lesion boundaries, reduces the loss of detailed features, enhances the stability and robustness of segmentation, and improves segmentation accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122289677A_ABST
    Figure CN122289677A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of artificial intelligence and medical image processing technology. It proposes a skin lesion image segmentation system and method based on a multi-scale information preservation mechanism. The system obtains segmentation results based on acquired skin lesion images and a preset target segmentation model. The target segmentation model is a cut neural network model, which divides the skin lesion image into multiple image blocks according to spatial location. A distance attenuation matrix for each attention head is constructed based on the distance between different image blocks. Different attenuation coefficients are set for different attention heads when constructing the distance attenuation matrix to preserve information at different spatial ranges. By introducing distance-based attenuation constraints into multi-head attention and setting differentiated attenuation coefficients for different attention heads, selective preservation and transmission of multi-scale information are achieved, reducing the loss of detailed features and improving the ability to characterize lesion boundaries.
Need to check novelty before this filing date? Find Prior Art