Computer assisted detection and characterisation of breast cancer in MRI

This thesis presents a novel set of CAD (computer assisted detection/diagnosis) tools to assist the radiologist with the increasingly complex task of detecting and characterising breast lesions using magnetic resonance imaging (MRI). Commercial CAD systems presently fall short of automatically locating and classifying malignant lesions. Instead they automate many of the image processing and analysis functions that would otherwise have to be performed manually and visualise the data to aid interpretation. It is perhaps not surprising, therefore, that a recent meta-study concluded that existing breast MRI CAD does not improve the sensitivity and specificity of experienced radiologists and their interpretation remains essential. A recent review of breast MRI and MR spectroscopy concluded that what is needed are “quantitative features extracted preferably from the automatically segmented 3D lesion” and a more comprehensive assessment of lesions based on features/measurements “derived from MR multi-parametric acquisitions”. This then was the motivation for the objectives of this thesis: (i) to develop an automatic 3D lesion segmentation algorithm for multi-modal breast MRI data; (ii) to develop features that quantitatively characterise the lesion (morphology, microvasculature, and microstructure) and other breast cancer signs from this data; and (iii) to evaluate these CAD tools using clinical breast MRI data. With regard to (i) a novel fully automatic method for segmentation (i.e., detection and delineation) of suspicious tissue in breast MRI is presented and evaluated. The method is based on mean-shift clustering and graph-cuts on a region adjacency graph. To the author's knowledge it is the first fully automatic method for breast lesion detection and delineation in breast MRI. The method was tested on a total of 102 lesions from two different vendors' scanner systems. The regions of interest identified by the method were compared with the ground truth (manually delineated by an experienced radiographer) and the detection and delineation accuracies quantitatively evaluated. One hundred percent of the lesions were detected with a mean of 4.5±1.2 false positives per subject. This false-positive rate is nearly 50% better than previously reported for a fully automatic breast lesion detection system. The median Dice coefficient was 0.76 (interquartile range, 0.17), and 0.75 (interquartile range, 0.16) for the two scanner systems respectively. With regard to (ii) several new features for breast MRI CAD—derived from anatomical T1w and T2w images, DCE-MRI, and DW-MRI—are presented. They include features that characterise vascularity, blooming, and centripetal/centrifugal enhancement; and features extracted from the repartition of a lesion into mean-shift clusters. A new method for the fully automatic segmentation and measurement of the internal mammary vessels is also presented and evaluated. This was motivated by recent findings that the vascular cross-sectional area of internal mammary vessels is significantly larger on the side of breast cancer compared to the contralateral side. The method was developed on clinical MR data from 15 subjects, and tested on a further 145. The results do not indicate any significant difference in cross-sectional area between subjects with malignant lesions and those without (p > 0.05). With regard to (iii) the efficacy of the proposed features was evaluated using both manual and automatic segmentations of 74 malignant and 47 benign lesions. A superset of proposed and state-of-the-art features was computed for each lesion. Random forest classification was used to estimate classification performance—area under the receiver operating characteristic curve (AUC)—and to identify the most important features. This was done independently for mass-like lesions, non-mass-like lesions, and a combination of both. The results show that the AUCs for the two segmentation approaches are on a par; the best AUC is for mass-like lesions: 0.874±0.038 (manual) and 0.873±0.039 (automatic); the worst AUC is for non-mass-like lesions: 0.677±0.081 and 0.663±0.082 respectively; and the proportion of new features in the top 10 features ranges from 20-60%. Results also show that the combination of T2w, DCE-MRI, and DW-MRI features yields the best performance; i.e. T1w features offer little if any improvement in performance. In summary the CAD tools developed in this thesis permit fully automatic detection and delineation of suspicious lesions, the extraction of lesion features—including several new features characterising vascularity, blooming, and centripetal/centrifugal enhancement—from multi-modal MR images, and the classification of lesions as benign or malignant. The experimental results show that classification performance based on automatically segmented lesions is as good as that for manually segmented lesions. This suggests that it is indeed possible for fully automatic CAD to achieve the sensitivity/specificity of an experienced radiologist.