Nephroblastoma in MRI Data

Abstract

The main objective of this work is the mathematical analysis of nephroblastoma in MRI sequences. At the beginning we provide two different datasets for segmentation and classification. Based on the first dataset, we analyze the current clinical practice regarding therapy planning on the basis of annotations of a single radiologist. We can show with our benchmark that this approach is not optimal and that there may be significant differences between human annotators and even radiologists. In addition, we demonstrate that the approximation of the tumor shape currently used is too coarse granular and thus prone to errors. We address this problem and develop a method for interactive segmentation that allows an intuitive and accurate annotation of the tumor. While the first part of this thesis is mainly concerned with the segmentation of Wilms’ tumors, the second part deals with the reliability of diagnosis and the planning of the course of therapy. The second data set we compiled allows us to develop a method that dramatically improves the differential diagnosis between nephroblastoma and its precursor lesion nephroblastomatosis. Finally, we can show that even the standard MRI modality for Wilms’ tumors is sufficient to estimate the developmental tendencies of nephroblastoma under chemotherapy.