Layer Embedding Analysis in Convolutional Neural Networks for Improved Probability Calibration and Classification.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Machine learning interpretability

Background:

• Interpreting deep learning models, especially in medical applications, remains a challenge.
• Understanding layer-by-layer decisions is crucial for trust and performance improvement.
• Existing methods often lack interpretability or efficiency.

Purpose of the Study:

• To develop a method for interpreting neural network layer-by-layer embedded decisions.
• To utilize these insights for enhancing model performance in classification tasks.
• To improve probability calibration and classification accuracy in medical imaging.

Main Methods:

• Approximating image representation distributions in embeddings using random forest models.
• Generating 'embedding outputs' to measure sample classification.
• Developing a pipeline to use embedding outputs for model calibration and performance enhancement.

Main Results:

• The method provides interpretable visualizations of neural network decisions.
• Embedding-derived information significantly improves probability calibration and classification accuracy.
• Achieved superior performance compared to baseline methods on liver tissue classification tasks (3D MR and CT images).

Conclusions:

• The proposed method offers an effective and computationally efficient approach to interpret neural networks.
• It enhances model performance through improved probability calibration and classification.
• The technique is robust, easy to implement, and valuable for medical image analysis.