Related Experiment Video
Updated: Jun 24, 2025
06:08
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
155
Expansive Receptive Field and Local Feature Extraction Network: Advancing Multiscale Feature Fusion for Breast
Yongxin Guo1,2, Yufeng Zhou3,4,5
1Medical College Road, State Key Laboratory of Ultrasound in Medicine and Engineering, College of Biomedical Engineering, Chongqing Medical University, Chongqing, 400016, China.
Journal of Imaging Informatics in Medicine
|May 31, 2024
Summary
A novel deep learning network accurately segments breast fibroadenomas in sonography. This computer-aided diagnosis (CAD) tool enhances early detection and improves diagnostic efficiency for this common benign breast disease.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Fibroadenoma is a prevalent benign breast condition impacting women across all age groups.
- Accurate and efficient diagnosis of breast fibroadenomas is crucial for effective treatment and pain reduction.
- Computer-aided diagnosis (CAD) shows promise for enhancing diagnostic accuracy and efficiency in medical imaging, yet its use in sonography remains limited.
Purpose of the Study:
- To develop and evaluate a deep learning network for accurate segmentation of breast fibroadenomas in ultrasound (sonography) images.
- To improve the application of computer-aided diagnosis in breast fibroadenoma detection using sonographic data.
Main Methods:
- Proposed a novel network architecture incorporating a Hierarchical Attentive Fusion module for local information learning (channel-wise and pixel-wise) and a Residual Large-Kernel module for global information learning via multiscale convolutions.
- Integrated multiscale feature fusion within both modules to enhance network stability.
- Employed an energy function and data augmentation techniques for fine-tuning low-level medical image features and improving data enhancement.
Main Results:
- Achieved Mean Pixel Accuracy (MPA) of 93.93% and Mean Intersection over Union (MIOU) of 88.16% on a local clinical dataset.
- Attained MPA of 86.06% and MIOU of 73.19% on a public dataset, significantly outperforming state-of-the-art methods like SegFormer.
- Demonstrated superior feature extraction capabilities by combining local pixel-wise learning with expansive receptive fields for global perception.
Conclusions:
- The proposed deep network effectively segments breast fibroadenomas in sonography, showcasing powerful local-global feature extraction.
- This advanced segmentation capability holds significant potential for improving the early diagnosis of breast fibroadenomas.
- The developed CAD approach offers a valuable tool for enhancing diagnostic accuracy and efficiency in breast ultrasound.
