Understanding Bias in Multispectral Autofluorescence Lifetime Imaging: Are Models Sensitive to Oral Location?

Area of Science:

• Medical Artificial Intelligence
• Oral Oncology
• Biomedical Imaging

Background:

• AI fairness is critical in healthcare, as model outcomes directly impact patient health.
• Common AI bias sources include demographics, but oral location is a unique bias factor in oral cancer diagnosis.
• Existing AI models show performance variability across different oral locations, necessitating further investigation.

Purpose of the Study:

• To investigate the impact of oral location as a source of bias in AI-driven oral cancer diagnosis.
• To understand how multispectral autofluorescence imaging is affected by oral location and lesion presence.
• To develop strategies for creating fairer and more equitable AI diagnostic models.

Main Methods:

• Designed three experiments to analyze the effect of oral location on AI model performance.
• Utilized multispectral autofluorescence images to assess tissue characteristics and disease-associated features.
• Proposed and evaluated a tissue-specific fine-tuning approach.

Main Results:

• Multispectral autofluorescence images capture tissue characteristics, but this information degrades in lesion images.
• Tissue-specific features were found to be intertwined with disease-associated features.
• The proposed tissue-specific fine-tuning approach improved overall performance and reduced the fairness gap by over 5%.

Conclusions:

• Oral location is a significant source of bias in AI models for oral cancer diagnosis.
• AI models require careful design to mitigate location-based biases and ensure equitable performance.
• Tissue-specific fine-tuning offers a promising method to enhance fairness and accuracy in oral cancer diagnostic AI, improving clinical outcomes.