M-VAAL: Multimodal Variational Adversarial Active Learning for Downstream Medical Image Analysis Tasks.

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• Acquiring annotated medical data is expensive due to expert involvement, time-consuming protocols, and validation requirements.
• Active learning aims to reduce the need for extensive annotated data by selecting the most informative samples for annotation.
• Active learning is crucial for improving deep learning models in medical diagnosis, assessment, and treatment planning.

Purpose of the Study:

• To propose a novel active learning method that leverages multimodal auxiliary information for enhanced sample selection.
• To address the limitations of existing task-specific active learning methods in medical image analysis.
• To improve the robustness and efficiency of data-driven medical AI models with limited annotations.

Main Methods:

• Developed a Multimodal Variational Adversarial Active Learning (M-VAAL) method.
• Integrated auxiliary information from additional modalities into the active learning sampler.
• Applied the M-VAAL method to brain tumor segmentation/classification (BraTS2018) and chest X-ray classification (COVID-QU-Ex).

Main Results:

• Demonstrated the effectiveness of M-VAAL in enhancing active sampling using multimodal data.
• Achieved promising results in data-efficient learning for medical image analysis tasks with limited annotations.
• Showcased the potential of M-VAAL for improving deep learning model performance in clinical applications.

Conclusions:

• The proposed M-VAAL method offers a robust and data-efficient approach for medical AI development.
• Leveraging multimodal information significantly enhances active learning sample selection.
• M-VAAL shows promise for reducing the cost and effort associated with medical data annotation.