A Diffusion-Assisted Attention U-Net++ Framework for Accurate Breast Tumor Segmentation in Ultrasound Imaging

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216305%3A26220%2F26%3A0199588" target="_blank" >RIV/00216305:26220/26:0199588 - isvavai.cz</a>

Výsledek na webu

<a href="https://www.spiedigitallibrary.org/conference-proceedings-of-spie/14017/140170E/A-diffusion-assisted-attention-U-Net-framework-for-accurate-breast/10.1117/12.3100593.full" target="_blank" >https://www.spiedigitallibrary.org/conference-proceedings-of-spie/14017/140170E/A-diffusion-assisted-attention-U-Net-framework-for-accurate-breast/10.1117/12.3100593.full</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1117/12.3100593" target="_blank" >10.1117/12.3100593</a>

Jazyk výsledku

angličtina

Název v původním jazyce

A Diffusion-Assisted Attention U-Net++ Framework for Accurate Breast Tumor Segmentation in Ultrasound Imaging

Popis výsledku v původním jazyce

Accurate segmentation of breast tumors in ultrasound (US) imaging is challenging due to speckle noise, low contrast, and irregular lesion morphology. This study proposes a diffusion-assisted Attention U-Net++ framework that integrates anisotropic diffusion filtering with nested attention skip pathways to enhance noise suppression and boundary precision. The anisotropic diffusion stage reduces speckle artifacts while preserving structural edges, and the Attention U-Net++ network refines multiscale feature fusion through attention gating. Experiments were conducted on the Breast Ultrasound Images (BUSI) dataset comprising 780 images from 600 subjects. The proposed framework achieved a Dice coefficient of 0.965, IoU of 0.932, precision of 0.97, recall of 0.98 for benign lesions, and precision of 0.965, recall of 0.94 for alignant lesions, with an overall accuracy of 0.95. Grad-CAM visualization confirmed that the model focused on clinically relevant tumor regions, enhancing interpretability. The results demonstrate that combining diffusion-based preprocessing with attention-guided segmentation yields robust, efficient, and explainable performance, offering a reliable foundation for future computer-aided breast-ultrasound diagnostic systems.

Název v anglickém jazyce

A Diffusion-Assisted Attention U-Net++ Framework for Accurate Breast Tumor Segmentation in Ultrasound Imaging

Popis výsledku anglicky

Accurate segmentation of breast tumors in ultrasound (US) imaging is challenging due to speckle noise, low contrast, and irregular lesion morphology. This study proposes a diffusion-assisted Attention U-Net++ framework that integrates anisotropic diffusion filtering with nested attention skip pathways to enhance noise suppression and boundary precision. The anisotropic diffusion stage reduces speckle artifacts while preserving structural edges, and the Attention U-Net++ network refines multiscale feature fusion through attention gating. Experiments were conducted on the Breast Ultrasound Images (BUSI) dataset comprising 780 images from 600 subjects. The proposed framework achieved a Dice coefficient of 0.965, IoU of 0.932, precision of 0.97, recall of 0.98 for benign lesions, and precision of 0.965, recall of 0.94 for alignant lesions, with an overall accuracy of 0.95. Grad-CAM visualization confirmed that the model focused on clinically relevant tumor regions, enhancing interpretability. The results demonstrate that combining diffusion-based preprocessing with attention-guided segmentation yields robust, efficient, and explainable performance, offering a reliable foundation for future computer-aided breast-ultrasound diagnostic systems.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

20601 - Medical engineering

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2026

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

PROCEEDINGS OF SPIE Ninth International Conference on Advances in Image Processing (ICAIP 2025)

ISBN

9781510699915

ISSN

—

e-ISSN

—

Počet stran výsledku

9

Strana od-do

—

Název nakladatele

SPIE

Místo vydání

Bellingham, Washington, USA

Místo konání akce

Chengdu, PCR

Datum konání akce

7. 11. 2025

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—