Related Experiment Videos
A metric learning-based method using graph neural network for pancreatic cystic neoplasm classification from CTs.
Jiachen Zhang1, Yishen Mao2, Ji Li2
1School of Information Science and Engineering, East China University of Science and Technology, Shanghai, China.
Medical Physics
|May 10, 2022
Summary
A novel graph neural network (GNN) approach accurately classifies pancreatic cystic neoplasms (PCNs) using computed tomography (CT) images, even with small, imbalanced datasets. This method improves diagnostic accuracy and reduces labeling costs for computer-aided diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Pancreatic cystic neoplasms (PCNs) present diagnostic challenges due to their rarity and difficulty in preoperative classification.
- Traditional deep learning models require large labeled datasets and precise segmentation, which are often unavailable for PCNs.
- Accurate classification is crucial for appropriate patient management and treatment planning.
Purpose of the Study:
- To develop a metric learning-based graph neural network (GNN) method for accurate and efficient classification of PCNs from computed tomography (CT) images.
- To address the limitations of small and imbalanced datasets in PCN classification.
- To improve preoperative diagnostic support for clinicians.
Main Methods:
- A framework utilizing a convolutional neural network (CNN) for feature extraction from CT images.
- Integration of a GNN with a metric learning strategy to model similarities between feature vectors and classify PCNs.
- Implementation of subtasks with randomly selected images to enhance model generalization.
Main Results:
- The GNN-based model achieved high performance on two classification tasks: benign/malignant diagnosis (88.926% accuracy) and specific type classification (74.497% accuracy).
- The method effectively mitigated the negative impact of imbalanced datasets, as evidenced by improved F1 scores and macroaverage comparisons.
- Demonstrated superior performance compared to existing methods, particularly on small and imbalanced datasets.
Conclusions:
- The proposed GNN-based approach offers a promising solution for PCN classification, outperforming existing models on challenging datasets.
- This method reduces the need for extensive labeling, thereby lowering costs associated with computer-aided diagnosis.
- The findings support the potential application of this GNN model in clinical settings for computer-aided diagnosis of PCNs.