A Question-Centric Model for Visual Question Answering in Medical Imaging.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning excels in medical image analysis but lacks transparency, raising concerns about clinical use.
• Current methods for understanding model behavior rely on indirect strategies like uncertainty estimation.
• Explicitly querying models offers a more direct approach to assessing their performance and failure modes.

Purpose of the Study:

• To develop a novel Visual Question Answering (VQA) approach for interpretable medical image analysis.
• To enable direct querying of deep learning models using natural language questions about image content.
• To enhance the transparency and trustworthiness of AI in clinical settings.

Main Methods:

• A novel Visual Question Answering (VQA) framework was developed.
• The approach fuses image and question features in a unique manner.
• Experiments were conducted on diverse medical and natural image datasets.

Main Results:

• The proposed VQA method achieved accuracy equal to or higher than existing approaches.
• The fusion of image and question features proved effective for model querying.
• Demonstrated the viability of direct querying for understanding model behavior.

Conclusions:

• Visual Question Answering presents a promising direct method for interpreting deep learning models in medical imaging.
• The novel VQA approach enhances transparency and can aid in clinical adoption of AI.
• Further research can build upon this method to improve AI explainability in healthcare.