Medical Tumor Image Classification Based on Few-Shot Learning.

Area of Science:

• Medical image analysis
• Machine learning in healthcare
• Computational pathology

Background:

• Cancer poses a significant health burden, necessitating accurate diagnostic tools.
• Pathologist-based assessment of pathological images is subjective and time-consuming.
• Computer-aided diagnosis (CAD) systems offer potential for improved accuracy and efficiency.
• Deep learning for CAD requires large labeled datasets, which are scarce in medical imaging.

Purpose of the Study:

• To develop an improved few-shot learning method for medical image recognition.
• To address the challenge of limited labeled data in training machine learning models for CAD.
• To enhance the utilization of limited feature information from few samples.

Main Methods:

• Proposed an improved few-shot learning algorithm tailored for medical image recognition.
• Incorporated a feature fusion strategy to maximize information from limited samples.
• Evaluated the model on BreakHis and skin lesion datasets.

Main Results:

• Achieved 91.22% classification accuracy on the BreakHis dataset with only 10 labeled samples.
• Attained 71.20% classification accuracy on the skin lesion dataset with 10 labeled samples.
• Demonstrated superior performance compared to other state-of-the-art methods in few-shot learning scenarios.

Conclusions:

• The proposed improved few-shot learning method effectively enhances medical image recognition accuracy with minimal labeled data.
• The feature fusion strategy is crucial for leveraging limited information in few-shot learning.
• This approach shows significant promise for advancing CAD systems, particularly in data-scarce domains.