Testing and improvement of a prototype application for lesion classification and data creation for computerized image analysis for lesion detection in multiple myeloma

点击翻译按钮获取中文摘要

Multiple Myeloma (MM) is a systemic tumor disease in the bone marrow, which affects the skeleton. It causes multiple tumors in bones, but also osteolytic lesions, resulting in diffuse or focal dissolution of bone tissue. These lesions cause painful bone destruction and fractures. Currently, neither overall tumor mass nor the degree of skeletal changes can be quantified. Therefor the segmentation of every single lesion is needed. The large amount of lesions and the examination of the whole skeleton impede manual segmentation of all lesions in the clinician's daily routine. Delineation of tumors and osteolytic lesions requires multiple modalities because lesions show different characteristics in different modalities. Only by using automatic methods this segmentation task can be performed in a reasonable way. Currently, no automatic solutions for this problem are available. Knowing overall tumor volume and quantification of skeletal status might lead to more accurate staging and therapy monitoring. An important step for this analysis is the segmentation of single bones. Segmentation of single bones is also important in other applications. For the analysis of new treatment strategies in MM, for the irradiation of bones e.g., the volume of the affected hematopoietic bone marrow is relevant. Therefor the segmentation of the bones is necessary. Furthermore, segmentation of single bones is highly relevant for other applications in oncology and radiotherapy, since bones give important information for motion analysis and motion compensation during therapy. In this project, the data basis is to be created for further automation of image analysis in MM. An already developed prototype for lesion classification for lesion segmentation based on multimodal features was exemplarily tested on 4 MM patients. Aim of the project is the analysis and improvement of this prototype application. The application is to be trained and tested based on a larger data sample. Therefor the creation of training data is needed, which requires manual segmentation of lesions and single bones in whole body images. Manual segmentation of lesions is essential since it is the base for a pattern recognition method that uses multimodal features and learns characteristics of lesions and healthy tissues and should later be able to detect lesions in new whole body images. Manual bone segmentations are used first, to limit the search region for the prototype application to bone and second, to provide a training data sample for the development of an automatic bone segmentation method.