A novel adaptive momentum method for medical image classification using convolutional neural network

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are crucial for AI-driven medical diagnosis, particularly in image classification.
• Momentum is essential in stochastic gradient optimization for CNN training, but traditional constant weighting requires complex hyperparameter tuning.
• Tuning momentum hyperparameters in CNNs can be computationally intensive and challenging.

Purpose of the Study:

• To propose a novel adaptive momentum method for faster and more stable convergence in CNN training.
• To eliminate the need for manual momentum hyperparameter optimization in deep learning models.
• To enhance the performance of AI models in medical image classification tasks.

Main Methods:

• Developed an adaptive momentum rate that adjusts based on epoch error changes, eliminating manual hyperparameter tuning.
• Evaluated the proposed method on diverse medical datasets: REMBRANDT Brain Cancer, NIH Chest X-ray, and COVID-19 CT scans.
• Compared the novel adaptive momentum optimizer against Stochastic Gradient Descent (SGD), Adam, and RMSprop.

Main Results:

• The adaptive momentum method reduced classification error from 6.12% to 5.44% when improving SGD performance.
• Achieved the lowest error rates and highest accuracy compared to other tested optimizers.
• Demonstrated superior performance with state-of-the-art CNN architectures, reaching 95% accuracy on the same datasets.

Conclusions:

• The proposed adaptive momentum significantly improves CNN convergence and classification accuracy in medical imaging.
• This method offers a more efficient and stable alternative to traditional momentum optimizers.
• The adaptive momentum approach enhances AI diagnostic capabilities by improving model performance and reducing classification errors.