Preminda Kessar

Does registration improve the performance of a computer aided diagnosis system for dynamic contrast-enhanced MR mammography?

This study investigated whether image registration improves the classification performance of a computer aided diagnosis (CAD) system for dynamic contrast-enhanced (DCE) MR mammography The CAD system that we developed included image registration, semi-automatic lesion segmentation, 3D image features extraction, and feature selection and combination by logistic regression analysis. The CAD system achieved a leave-one-out area under the ROC curve of 0.86, which is within the range of reported classification performances. This performance was not the artifact of the feature selection process or the leave-one-out test procedure. Worse results were obtained without segmentation refinement and image registration. Rigid image registration led to a statistically significant increase of the area under the ROC curve from 0.81 to 0.86

Classification Improvement by Segmentation Refinement: Application to Contrast-Enhanced MR-Mammography

In this study we investigated whether automatic refinement of manually segmented MR breast lesions improves the discrimination of benign and malignant breast lesions. A constrained maximum a-posteriori scheme was employed to extract the most probable lesion for a user-provided coarse manual segmentation. Standard shape, texture and contrast enhancement features were derived from both the manual and the refined segmentations for 10 benign and 16 malignant lesions and their discrimination ability was compared. The refined segmentations were more consistent than the manual segmentations from a radiologist and a non-expert. The automatic refinement was robust to inaccuracies of the manual segmentation. Classification accuracy improved on average from 69% to 82% after segmentation refinement.

SOM-Based wavelet filtering for the exploration of medical images

In medical image analysis there are many applications that require the definition of characteristic image features. Especially computationally generated characteristic image features have potential for the exploration of large datasets. In this work, we propose a method for investigating time series of medical images using a combination of the Discrete Wavelet Transform and the Self Organizing Map. Our approach allows relevant image information to be identified in wavelet space. This enables us to develop a filter algorithm suitable to find and extract the characteristic image features and to suppress interfering non-relevant image information.

Content Based Image Retrieval for Dynamic Time Series Data

Content based image retrieval (CBIR) systems in the field of medical image analysis are an active field of research. They allow the user to compare a given case with others in order to assist in the diagnostic process. In this work a CBIR system is described working on datasets which are both time- and space-dependent. Different possible feature sets are investigated, in order to explore how these datasets are optimally represented in the corresponding database.