Related Experiment Video
Updated: Jul 10, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Densely connected convolutional networks for ultrasound image based lesion segmentation
Jinlian Ma1, Dexing Kong2, Fa Wu2
1School of Integrated Circuits, Shandong University, Jinan 250101, China; Shenzhen Research Institute of Shandong University, A301 Virtual University Park in South District of Shenzhen, China; State Key Lab of CAD&CG, College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China.
This study introduces MDenseNet, a deep learning model for accurate segmentation of thyroid and breast nodules in ultrasound images. The method improves diagnostic accuracy and reproducibility in clinical practice.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Accurate delineation of thyroid and breast lesions is crucial for cancer diagnosis, treatment planning, and outcome evaluation.
- Manual annotation of lesions in low-quality ultrasound (US) images is challenging due to noise, variable appearances, and ambiguous boundaries, leading to time-consuming and error-prone processes.
- Automated segmentation of nodular lesions in US images is highly desirable but technically difficult.
Purpose of the Study:
- To develop and evaluate a novel densely connected convolutional network (MDenseNet) for automated segmentation of nodular lesions in 2D ultrasound images.
- To enhance the MDenseNet architecture with pre-training strategies (PMDenseNet and PDMDenseNet) for improved performance in segmenting thyroid and breast nodules.
- To compare the proposed MDenseNet-based method against existing state-of-the-art convolutional neural networks.
Main Methods:
- A novel densely connected convolutional network (MDenseNet) was proposed for nodular lesion segmentation.
- The network was pre-trained on the ImageNet database (PMDenseNet) and subsequently fine-tuned on ultrasound datasets.
- A deeper version, PDMDenseNet, was designed by incorporating an additional dense block for enhanced segmentation of thyroid and breast nodules.
Main Results:
- The MDenseNet-based method demonstrated accurate segmentation of multiple nodular lesions, including those with complex shapes, in thyroid and breast ultrasound images.
- The proposed method outperformed three state-of-the-art convolutional neural networks in terms of segmentation accuracy and reproducibility.
- The study achieved promising results in segmenting nodular lesions from ultrasound images, indicating broad applicability.
Conclusions:
- The developed MDenseNet architecture with pre-training strategies offers an accurate and reproducible solution for automated nodular lesion segmentation in ultrasound imaging.
- The method shows significant potential for clinical application in diagnosing thyroid and breast cancers and can be extended to other medical segmentation tasks.
- This deep learning approach addresses the limitations of manual annotation, paving the way for more efficient and reliable diagnostic workflows.
