Clustering methods applied to diagnostic ultrasound of adnexal masses

The aim of the thesis is to explore problems related to ultrasound imaging, in particular to ovarian cancer, and to develop tools to analyze the images using unsupervised learning algorithm. Benign and malignant masses show different features that can be investigated with ultrasound scan, making this imaging technique an important and low cost instrument to gain relevant information about the potential tumor. Unfortunately the efficacy of this imaging technique is very much operator-dependent and the capability of discriminating functional ovarian masses from malignant ovarian tumor depends a lot on the doctor’s experience. Therefore computer-aided examination could help to reduce the gap in performance between doctors having different level of experience. During this project I focused mainly on two different tasks. I’ve analyzed different quantities that allow spectral clustering to distinguish between images of serous and mucinous cysts. Doing so I’ve worked with images cropped using three different techniques: using a window common to all the images, using a window adjusted for every image and using an automatic segmentation routine based on local entropy to find the region of interest. I’ve also applied clustering algorithms, such as spectral and agglomerative hierarchical clustering, to histological reports to find the words that trigger the classification of these documents into different classes. This is a first step towards a more complete set of algorithms aimed at helping the medical doctor in making a correct risk assessment of as many as possible type of ovarian cysts. These results could lead to the basis for an engine for contentbased retrieval of ovarian ultrasound images for assisting doctors. Future improvements need to use both supervised and unsupervised algorithms to automatically estimate the risk associated with every image and combine this information in a label for the whole ultrasound examination.