Work

DMSA-KANet: Improving Breast Cancer Segmentation in Mammography through Deep Supervision Multi-scale Attention and Kolmogorov-Arnold Networks

Year: 2026
Type: article
Abstract: Breast cancer segmentation in mammography remains challenging owing to the small lesion size, low contrast, and overlapping fibroglandular tissue.This study proposes the Deep Supervision Multi-Scale Attention Kolmogorov-Arnold Network (DMSA-KANet), a hybrid architecture that combines an EfficientNetB4 encoder with a high-resolution path, a multi-scale attention bottleneck refined sequentially by Kolmogorov-Arnold Network (KAN) layers, a Vision Transformer, and a KAN-based Convolutional Block Attention Module (KAN-CBAM), and a decoder with KAN-gated skip connections and deep supervision.The model was evaluated on the low-energy images of CDD-CESM (650 images, 282 patients) and on the mass subset of CBIS-DDSM (1,592 images, 892 patients) with patientlevel five-fold cross-validation, in which checkpoint and threshold selection used only validation patients drawn from the training folds, and it was compared with baselines retrained under the same folds, loss, augmentation, epoch budget, and selection rules.Under this protocol DMSA-KANet did not outperform the baselines: it achieved a Dice of 43.97 ± 9.11% on CDD-CESM, the lowest of the seven models, compared with 51.02 ± 3.94% for U-Net and 51.50 ± 3.37% for TransUNet, and 39.98 ± 8.40% on CBIS-DDSM (44.93 ± 1.96% with its learning rates re-selected within the training patients of each fold), compared with 46.80 ± 5.86% for U-Net and 51.06 ± 4.26% for TransUNet, with larger variation across folds than any successfully trained baseline.The gap was associated with fold-level over-segmentation failures and unstable decision thresholds.These results describe retrospective technical performance on two public datasets and indicate that the added architectural complexity did not yield more accurate lesion segmentation under patient-level evaluation.The study provides a reproducible patient-level benchmark for CDD-CESM and CBIS-DDSM, with all fold assignments, code, and per-fold results publicly released.
Language: English
Open Access status: hybrid
OpenAlex ID: W7235806718
FWCI:
Cites: 21
Cited by:
Related to: 10