研究目的
To propose a Deep Feature Aggregation network (DFA-Net) for main aortic segmentation from CTA by aggregating features from forwarding layers to leverage more visual information and achieve state-of-the-art performance.
研究成果
The proposed DFA-Net effectively segments the main aortic from CTA images by aggregating features from multiple layers, achieving a high mIoU of 0.943. The semi-automatic annotation algorithm significantly reduces annotation costs. This approach demonstrates state-of-the-art performance and suggests that feature concatenation is superior to summation for such tasks. Future work could involve larger datasets and further optimization of the network architecture.
研究不足
The method relies on manual pre-processing and seeding for the level-set annotation, which could introduce bias. The dataset is limited to 90 volumes from a single hospital, potentially affecting generalizability. Computational resources require multiple GPUs, which may not be accessible to all users.
1：Experimental Design and Method Selection:
The study uses a semi-automatic level-set based algorithm for data annotation and a deep feature aggregation network (DFA-Net) inspired by FCN and DenseNet for segmentation. The backbone is VGG net, with feature concatenation and bottleneck operations to aggregate features from different layers.
2：Sample Selection and Data Sources:
90 CTA volumes collected from Beijing AnZhen Hospital, comprising over 60,000 2-D slices. The dataset is divided into 70 instances for training, 5 for validation, and 15 for testing.
3：List of Experimental Equipment and Materials:
Matlab for annotation, PyTorch for deep learning, GPUs (Tesla K80) for training, and a Win7 platform with i7 core for annotation.
4：Experimental Procedures and Operational Workflow:
First, use the level-set algorithm to generate ground truth labels with manual seeding and pre-processing. Then, train DFA-Net using cross-entropy loss, mini-batch gradient descent with batch size 4 per GPU, momentum 0.99, and learning rate scheduling. Validation and testing follow to assess performance.
5：99, and learning rate scheduling. Validation and testing follow to assess performance.
Data Analysis Methods:
5. Data Analysis Methods: Performance measured using F-score, precision, recall, and mean Intersection-over-Union (mIoU) for segmentation accuracy.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
