Feature Extraction from Medical Images for an Oral Cancer Reoccurrence Prediction Environment

Abstract

We present the concept of a novel image feature extraction approach that will be used to predict oral cancer reoccurrence in the scope of the NeoMark project. Based on current clinical practice, we propose several numeric image features that characterize tumors and lymph nodes. In order to (semi) automatically extract those features we introduce the following approach which is independent from human subjectivity: Registration and supervised segmentation of CT/MR images forms the base of the automated extraction of geometric and texture features of tumors and lymph nodes. In order to reduce the amount of user interaction during follow ups we incorporate the segmentation results of the previous examinations. The robustness and the numeric manner of the extracted features make them ideally suited as input for a sophisticated adaptive prediction environment that estimates the likelihood of oral cancer reoccurrence and assists the clinician to develop a treatment plan.