Distance-weighted Sinkhorn loss for Alzheimer's disease classification

Area of Science:

• Machine Learning
• Biomedical Informatics
• Data Science

Background:

• Traditional loss functions inadequately penalize misclassified samples based on their feature space distance.
• A larger distance from the ground truth distribution suggests a higher penalty should be applied.

Purpose of the Study:

• To propose and evaluate a novel distance-weighted Sinkhorn (DWS) loss function.
• To improve the accuracy of neural network classification by incorporating sample distance into loss calculation.

Main Methods:

• Developed the distance-weighted Sinkhorn (DWS) loss framework.
• Applied the DWS framework with a neural network for Alzheimer's disease stage classification.
• Compared DWS performance against traditional neural network loss functions and machine learning methods.

Main Results:

• The DWS framework demonstrated superior performance compared to traditional neural network loss functions.
• DWS achieved comparable or better results than traditional machine learning methods in Alzheimer's disease classification.
• Empirical results highlight the effectiveness of DWS in handling feature space distances.

Conclusions:

• The proposed DWS loss offers a more effective penalty mechanism for misclassified samples.
• DWS shows significant potential for applications in biomedical informatics and data science, especially for medical image analysis.
• This approach enhances classification accuracy by considering the geometric relationship between samples in the feature space.