MADS-Net: a Multi-scale Attention and Dynamic Sparse Mask Fusion Network for Electrical Impedance Tomography Image Reconstruction in Breast Screening
10.3724/j.pibb.2026.0197CSTR：32369.14.pibb.20260197
- VernacularTitle:MADS-Net：面向乳腺筛查电阻抗断层成像重建的多尺度注意力与动态稀疏掩码融合网络
- Author:
Cheng-Tao YU
1
;
Jia-Hui WANG
1
;
Song-Pei HU
2
;
Huai-Jin CHEN
3
;
Li DING
1
;
Xia YE
1
;
Jia-Feng YAO
4
Author Information
1. College of Mechanical Engineering, Jiangsu University of Technology, Changzhou 213001, China
2. College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China
3. School of Future Technology, South China University of Technology, Guangzhou 510641, China
4. College of Physics and Optoelectronic Engineering, Jinan University, Guangzhou 510632, China
- Publication Type:Journal Article
- Keywords:
electrical impedance tomography;
dynamic mask network;
multi-scale attention mechanism
- From:
Progress in Biochemistry and Biophysics
2026;53(9):2480-2496
- CountryChina
- Language:Chinese
-
Abstract:
ObjectiveTo address the key challenges in breast electrical impedance tomography (EIT) reconstruction including the severe ill‑posedness of the inverse problem, limited measurement information from the electrodes, low contrast of small lesions, as well as the common drawbacks of traditional reconstruction methods such as boundary blurring, low spatial resolution, and weak small-target detection capability, and to overcome the limitations of existing deep learning approaches namely low training efficiency, insufficient physical consistency, and substantial information loss during the low‑to‑high dimensional mapping process, this study proposes a multi‑scale attention and dynamic sparse mask fusion network (MADS‑Net). The proposed method aims to enhance high‑resolution reconstruction accuracy, small-lesion detectability, noise robustness, and computational efficiency in breast‑screening-oriented EIT, thereby providing a feasible solution for the application of EIT in breast functional imaging. MethodsThe proposed network adopts a strategy of “locate first, reconstruct later”. It realizes adaptive allocation of computational resources through a dynamic mask module, alleviates information loss in the low-to-high dimensional mapping process via a multi-scale attention mechanism, and embeds physical constraints to improve the physical rationality of reconstruction results. The performance of the proposed method was validated through systematic numerical simulations, agar phantom experiments, and porcine tissue experiments. Comparisons were made with mainstream methods including the Gauss-Newton algorithm, convolutional neural networks, U-Net, and Deep D-Bar. The image correlation coefficient (ICC), root mean square error (RMSE), and peak signal-to-noise ratio (PSNR) were adopted as quantitative evaluation metrics. Furthermore, lesion‑level evaluation metrics were introduced to assess the capability of different methods in reconstructing abnormal target boundaries and small‑sized target regions. ResultsExperimental results demonstrate that MADS-Net outperforms all comparison methods across all test scenarios. In numerical simulations, it achieves higher accuracy in target morphology and spatial location recovery as well as better background suppression. In agar phantom experiments, its detection and imaging reliability for small targets is significantly superior to the comparison methods. In porcine tissue experiments, the model maintains good generalization performance and can adapt to complex biological tissue imaging scenarios. Quantitative results show that MADS-Net achieves an average ICC of 0.917 6±0.048 8, an RMSE as low as0.060 6±0.031 6, and a PSNR of (23.893 9±2.685 3) dB. It successfully detects tiny targets as small as 5 mm within a 100 mm diameter sensing field, reduces the total training duration by 23.5% compared with the baseline method, and maintains stable reconstruction performance even under strong noise interference. ConclusionThe results demonstrate that MADS-Net achieves favorable EIT reconstruction performance in numerical simulations, agar phantom experiments, and porcine tissue experiments, even under common practical imaging challenges including low signal-to-noise ratio and blurred tissue boundaries. This study provides a methodological reference for further research on EIT in breast-screening-related scenarios and other medical functional imaging tasks, and also offers a low-cost, radiation-free alternative for breast lesion screening that aligns with the clinical demand for non-invasive diagnostic tools. It should be noted that real clinical breast data have not yet been used for validation in this study, and the clinical applicability of the proposed method requires further investigation. Future work will prioritize collecting multi-center real clinical breast datasets, and further optimize the model’s generalization ability to adapt to individual differences in breast tissue composition to promote its clinical application.
