Diversity Learning Based on Multi-Latent Space for Medical Image Visual Question Generation.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Clinical Decision Support

Background:

• Auxiliary clinical diagnosis aims to address resource disparities but remains physician-dependent.
• Current intelligent systems have limited involvement in the diagnostic process.
• Existing medical question generation lacks diversity, hindering comprehensive patient evaluation.

Purpose of the Study:

• To develop a diversity learning-based visual question generation model for automated clinical diagnosis.
• To enhance patient information capture from multiple perspectives with reduced physician involvement.
• To generate informative and diverse question sets from medical images for interactive diagnosis.

Main Methods:

• Proposed a diversity learning-based visual question generation model utilizing a multi-latent space.
• Embedded visual and language information in different latent spaces to generate diverse questions.
• Introduced a novel loss function for diversity training and a new metric (similarity) for evaluation.
• Enabled control over generated question categories for directional questioning.

Main Results:

• The model successfully generated diverse and informative questions from medical images on Slake and VQA-RAD datasets.
• Demonstrated the effectiveness of the multi-latent space and diversity loss in enhancing question variety.
• The proposed similarity metric accurately evaluated model performance.

Conclusions:

• The developed model can generate diverse questions, improving automated clinical diagnosis.
• This approach supports interactive automated clinical diagnosis by working with answering models.
• The method offers a cost-effective solution for generating diagnostic datasets, reducing manual annotation efforts.