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

While bias in artificial intelligence is gaining attention across applications, model fairness is especially concerning in medical applications because a person's health may depend on the model outcome. Sources of bias in medical applications include age, gender, race, and social history. However, in oral cancer diagnosis, the oral location may be a source of bias. Variability in performance based on the oral location has been reported but is not well understood. To help ensure that models perform equitably regardless of location, we design three experiments to study the effect of oral location on model performance. We show that multispectral autofluorescence images retain tissue-type characteristics, but that the tissue-specific information is degraded in lesion images. Furthermore, we show that the tissue-specific features are not disentangled from the disease-associated features. Our results show that automated diagnosis models need to be thoughtfully designed to remove bias from the oral location to ensure equitable performance. Based on these insights, we propose a tissue-specific fine-tuning approach that increases overall performance and lowers the fairness gap by over 5%.Clinical relevance- This paper explores sources of offtarget variance in multispectral autofluorescence images. By understanding sources of bias in multispectral autofluorescence images, fairer and more robust models for oral cancer diagnosis and margin delineation can be developed, leading to greater clinical acceptance and more equitable patient outcomes.