SECNet: a self-evolving attention network with uncertainty-driven residual refinement for breast ultrasound cancer classification

Breast cancer diagnosis from ultrasound imaging remains a challenging task due to low image contrast, speckle noise, heterogeneous tissue appearance, and significant overlap between benign and malignant lesion characteristics. These factors often lead to diagnostic uncertainty and limit the reliability of conventional deep learning models that rely on static feature representations and fixed attention mechanisms. Improving both classification accuracy and prediction reliability is therefore essential for the development of clinically applicable computer-aided diagnosis systems. In this work, we propose SECNet, a self-evolving cross-domain cancer intelligence network that explicitly addresses diagnostic ambiguity by jointly modeling dynamic attention evolution, uncertainty estimation, and uncertainty-guided residual refinement within a unified end-to-end framework. a unified end-to-end framework. SECNet integrates an Adaptive Attention Evolution Unit that persistently recalibrates both spatial and channel-wise attention throughout training iterations, allowing the model to adapt to varying feature saliency in diverse imaging contexts. Additionally, an uncertainty-driven residual refinement module selectively modifies unstable feature regions based on assessed epistemic uncertainty, thereby enhancing robustness and mitigating overconfident predictions. The suggested technique is assessed using a publicly accessible breast ultrasonography dataset using a stratified and reproducible experimental design. Performance is evaluated by various clinically pertinent parameters, including accuracy, sensitivity, specificity, Cohen’s kappa, Micro-F1 score, AUROC, and AUPRC. SECNet attains an overall classification accuracy of 91.8 ± 0.7%, a Micro-F1 score of 91.4 ± 0.8%, and an average AUROC of 98.5%, achieving the best average performance among the evaluated convolutional, transformer-based, and ensemble benchmarks. Statistical analysis further confirms that the observed performance differences are significant for several baseline comparisons, while SECNet remains statistically comparable with some strong ensemble-based approaches. These results demonstrate that integrating attention evolution with uncertainty-aware residual refinement leads to more reliable and interpretable breast ultrasound cancer classification, highlighting SECNet’s potential for real-world clinical decision support.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-19
DOI
https://doi.org/10.1038/s41598-026-80292-9
Primary Topic
AI in cancer detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

SECNet: a self-evolving attention network with uncertainty-driven residual refinement for breast ultrasound cancer classification

Fadhil Salman Abed, Soran Abdulkarim Pasha
Scientific Reports
AI in cancer detection
article

SECNet: a self-evolving attention network with uncertainty-driven residual refinement for breast ultrasound cancer classification

Fadhil Salman Abed, Soran Abdulkarim Pasha
article en

Abstract

Breast cancer diagnosis from ultrasound imaging remains a challenging task due to low image contrast, speckle noise, heterogeneous tissue appearance, and significant overlap between benign and malignant lesion characteristics. These factors often lead to diagnostic uncertainty and limit the reliability of conventional deep learning models that rely on static feature representations and fixed attention mechanisms. Improving both classification accuracy and prediction reliability is therefore essential for the development of clinically applicable computer-aided diagnosis systems. In this work, we propose SECNet, a self-evolving cross-domain cancer intelligence network that explicitly addresses diagnostic ambiguity by jointly modeling dynamic attention evolution, uncertainty estimation, and uncertainty-guided residual refinement within a unified end-to-end framework. a unified end-to-end framework. SECNet integrates an Adaptive Attention Evolution Unit that persistently recalibrates both spatial and channel-wise attention throughout training iterations, allowing the model to adapt to varying feature saliency in diverse imaging contexts. Additionally, an uncertainty-driven residual refinement module selectively modifies unstable feature regions based on assessed epistemic uncertainty, thereby enhancing robustness and mitigating overconfident predictions. The suggested technique is assessed using a publicly accessible breast ultrasonography dataset using a stratified and reproducible experimental design. Performance is evaluated by various clinically pertinent parameters, including accuracy, sensitivity, specificity, Cohen’s kappa, Micro-F1 score, AUROC, and AUPRC. SECNet attains an overall classification accuracy of 91.8 ± 0.7%, a Micro-F1 score of 91.4 ± 0.8%, and an average AUROC of 98.5%, achieving the best average performance among the evaluated convolutional, transformer-based, and ensemble benchmarks. Statistical analysis further confirms that the observed performance differences are significant for several baseline comparisons, while SECNet remains statistically comparable with some strong ensemble-based approaches. These results demonstrate that integrating attention evolution with uncertainty-aware residual refinement leads to more reliable and interpretable breast ultrasound cancer classification, highlighting SECNet’s potential for real-world clinical decision support.

Scientific Reports
Sulaimani Polytechnic University (IQ), University of Garmian (IQ)
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon