A deep image classification model based on prior feature knowledge embedding and application in medical diagnosis

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Image classification faces challenges with subtle features, high target correlation, and low signal-to-noise ratios.
• Existing methods struggle to effectively embed prior feature knowledge, limiting performance in complex datasets.

Purpose of the Study:

• To propose a novel deep learning model for enhanced image classification, especially for medical images.
• To integrate prior feature knowledge effectively into the classification process.
• To improve accuracy and robustness in challenging image classification tasks.

Main Methods:

• A hybrid model using ResNet50 for visual feature extraction (feature pyramid, self-attention) and Radial Basis Probabilistic Neural Network (RBPNN) for prior knowledge embedding.
• Utilized channel cosine similarity attention and dynamic C-means clustering to select representative sample features for RBPNN kernel centers.
• Implemented a pattern aggregation layer in RBPNN to combine subcategory features into category features for final classification via Softmax.

Main Results:

• Achieved 85.82% accuracy on a brain tumor MRI dataset and 83.92% on a cardiac ultrasound dataset.
• Demonstrated significant improvements in performance indicators compared to three mainstream image classification models.
• The proposed method effectively highlights local and structural features, forming non-convex decision boundaries and reducing sample set completeness requirements.

Conclusions:

• The proposed ResNet-RBPNN deep learning model offers a robust solution for image classification, particularly in medical imaging.
• Effective embedding of prior feature knowledge and multi-scale feature extraction contribute to superior performance.
• The method shows promise for applications requiring high accuracy with complex or noisy image data.