UIUC Quant Brownbag

Human and AI melanoma detection

Joseph Houpt - University of Texas at San Antonio

Melanoma is highly prevalent in the general population, with nearly 1 in 28 people being diagnosed in their lifetime and rates continuing to rise.  Fortunately, early detection has a large effect on survival rates.  Initial detection of potential melanoma often relies on perceptual expertise, whether in support of the ABCDE rule in use by the general population or in visual assessment by an expert dermatologist.  Many have suggested that computer vision will become central to early detection and potentially even remove reliance on human vision.  I will give an overview of our research on how novices perceive skin lesions when focused on a subset of the ABCDE diagnostic dimensions (asymmetry, boarder irregularity, and color variation).  In one study, I demonstrate that people perceive valuable information for identifying melanoma beyond that which is extracted by state-of-the-art computer vision algorithms. In the next study, I will present an examination of the ABC criteria using General Recognition Theory. Finally, I will demonstrate the effect of perceptual expertise training, whether on ABC features individually, or on higher order features extracted from deep convolutional neural networks, influences skin lesion perception.  I will then highlight some cautions for using black-box computer vision systems for diagnosis based on our human perceptual training research.