Master thesis presentation – Supervoxel based algorithms for use in Breast MRI CAD Systems
Presented by Magnus Ziegler
Breast Cancer (BCa) is the most common cancer for women in the Western world, and the American National Institutes of Health estimates that approximately 12.3 % of American women will be diagnosed with breast cancer at some point in their lifetime. Breast MRI is increasingly used clinically as an adjunct to traditional x-ray mammography and ultrasound for the detection and characterization of BCa. The interpretation of the large volume of image data acquired in a breast MRI exam is both a complex and time consuming task for the radiologist. Moreover, while the sensitivity of breast MRI to BCa is very high, its specificity is poor to moderate which results in large numbers of false positives. Computer automated detection/diagnosis (CAD) systems have been developed to address these issues. However, a recent meta-study of breast MRI CAD systems concluded that they have little effect on the sensitivity and specificity of experienced radiologists. Current literature suggests that performance gains may be achieved through 3D segmentations of suspicious lesions and the use of features (measurements) derived from multimodal MRI.
Typically a clinical breast MRI examination includes the acquisition of anatomical T1- and T2-weighted images, and a dynamic contrast-enhanced (DCE) sequence. In this work, a state-of-the-art method for creating spatially coherent clusters of similarly enhancing voxels, or voxels with similar signal intensity characteristics, was implemented and explored for use in a breast MRI CAD system. The method, Simple Linear Iterative Clustering (SLIC), generates an oversegmentation of the image into regions termed superpixels (2D) or supervoxels (3D). This method is an adapted k-means clustering approach that quickly and efficiently generates supervoxels/pixels by bounding the search area for similar voxels/pixels, and through the use of a simple distance metric. Here SLIC was used: i) to develop an algorithm to segment the breast-air boundary; and ii) to partition volumes of interest, corresponding to mass-like lesions, into supervoxels from which quantitative features describing the lesion are extracted. These features describe the contrast enhancement characteristics and diffusion characteristics of the lesion. Clinically acquired MRI data was used to evaluate both the segmentation method and the efficacy of the proposed features for discriminating between benign and malignant lesions.
Segmentations of the breast-air boundary were reviewed visually and found to adhere well to the boundary. Random Forest classification was used to estimate the classification performance of the proposed features, as well as to identify the most important subset of features. The results, based on a study of 77 subjects, show that the classifier is able to discriminate between benign and malignant lesions with an accuracy of 0.752 ± 0.055 (AUC±SE). Collectively the results provide evidence that SLIC generated supervoxels are useful for both segmentation and classification in a CAD system. Further research is needed to investigate whether the combination of the proposed SLIC-based features and conventional features can improve the state-of-the-art in terms of sensitivity and specificity.
Title: Supervoxel based algorithms for use in Breast MRI CAD Systems
Date: Wednesday June 11th, 1400-1500
Location: Blå rummet, Chalmers S2
Thesis Examiner: Andrew Mehnert