Designing AI Components for Diagnostics of Carotid Body Tumors

Carotid body tumor (chemodectoma) is a rare disease, characterized by a large number of diagnosis errors and a lack of specialized software for its computational detection. Apparently, any datasets, representing this kind of tumor, do not exist. Thus, there is a problem of awareness among specialists. In this paper, we provide our own open access Python library for medical image preprocessing and tumor detection RadImaLib, which is intended to be useful in terms of the issues mentioned above. Having explored other similar libraries and several studies on the topic, we have formed and integrate into the library a dataset of chemodectoma images, made distinct methods for dataset creation for encouraging users to manipulate with their own medical data, implemented some image preprocessing methods, such as erosion, dilation, Hounsfield Scale transformation and rescaling, and developed our own detection model, based on the U-Net neural network architecture with a descent result accuracy.