Robust adversarial uncertainty quantification for deep learning fine-tuning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep learning models often struggle with uncertain or unpredictable inputs.
• Accurate medical image analysis is crucial for timely and correct diagnoses.
• Existing models may require transfer learning, increasing complexity.

Purpose of the Study:

• To develop a deep learning model robust to highly uncertain inputs.
• To enhance the performance and accuracy of machine learning models in image classification.
• To reduce the risk of misdiagnosis in medical imaging applications.

Main Methods:

• A three-phase approach: dataset creation, neural network design, and retraining for unpredictable inputs.
• Utilizing entropy values and a non-dominant sorting algorithm to identify high-entropy data points.
• Merging training sets with adversarial samples and updating dense network parameters using mini-batches.

Main Results:

• The proposed model demonstrated improved accuracy on the MNIST dataset (0.85 to 0.88) and the COVID dataset (0.83 to 0.85).
• Successful image classification was achieved for both datasets without employing transfer learning.
• The method shows potential for improving categorization of radiographic images and diagnostic accuracy.

Conclusions:

• The developed deep learning model effectively handles uncertain inputs and enhances image classification accuracy.
• This approach offers a promising alternative for medical imaging analysis, potentially reducing misdiagnosis risks.
• The model's efficacy was validated on benchmark and medical datasets, highlighting its robustness and efficiency.