Purpose: Explore the possible use of mathematical observer models in dermoscopy.
Methods: Mathematical Observers (MOs) have been used extensively in a Virtual Clinical Trial context as surrogates for human observers. They have commonly been used for radiology imaging to model lesion detection tasks, for e.g. in mammography and lung cancer screening. Increasingly reliable MOs have been researched that account for limitations of the human visual system (contrast sensitivity function and dynamic adaptation of the eye) as well as complex lesion characteristics and shapes (spiculations).State-of-the-art MOs, commonly used for radiology, have been applied to dermoscopy images and their usefulness and limitations explored.
Results: Applying existing MO models, such as channelized hotelling observers, to dermoscopy is not straightforward and does not provide reliable results. Compared to radiology, dermoscopy images are color images so lesions are characterized by changes in color as opposed to intensity (as is the case for most state-of-the-art MO models). Little research has been performed on color MO models. In addition, skin lesions are complex and contain a wide range of features (e.g. globules, white veil, etc). An approach where a MO is trained to detect a single feature is not feasible.Instead, a novel architecture for a layered observer model is proposed, which contains a set of observer models, each targeting a feature/set of features. The individual observer outputs are then combined into an overall observer output.This can be applied to dermoscopy where human observer decisions are typically not based on the presence of individual lesion features, but rather by looking at overall lesion characteristics and presence or absence of groups of features.
Conclusion: In this work we outline prior MO model work and explore usefulness and limitations for use in dermoscopy imaging, while exploring adaptations for making MOs better suitable for use in dermoscopy imaging.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by the Flanders Agency for Innovation & Entrepreneurship in the context of the Baekeland research Grant - Virtual Clinical Trials for Dermatology, HBC.2017.0583
Not Applicable / None Entered.