Automatic Detection of 3D Breast Lesions in Dynamic Contrast-Enhanced MR Images

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is being increasingly used in the clinic as an adjunct to x-ray mammography and ultrasound for the detection and characterisation of breast cancer. The technique involves acquiring T1-weighted volume images of one or both breasts before and at several time points after the injection of a contrast agent. Interpretation of this 4D data is a complex task for the radiologist and is becoming more so with developments of higher fieldstrength MRI scanners and associated increases in spatial, temporal and contrast information. Several commercial computer-assisted detection/diagnosis (CAD) systems for breast MRI have been developed in recent years to help radiologists with this task. However, these systems at present fall short of automatically locating and classifying malignant lesions. It is not surprising, therefore, that the efficacy of breast MRI CAD remains an open question. A recent review concluded that breast MRI CAD needs to be based on “quantitative features extracted preferably from the automatically segmented 3D lesion”. This thesis deals specifically with the problem of automatically segmenting (i.e. delineating) 3D lesions in breast DCE-MRI data. In particular it reviews existing approaches and proffers a novel approach based on voxel-wise classification. The new approach involves assigning a “suspiciousness” score to each voxel using features extracted from its time series, and then computing the spatial co-occurrence of this score in a neighbourhood including the voxel. The thesis also presents an empirical evaluation of the efficacy of this technique versus a competing method based on multispectral co-occurrence. The evaluation was performed on real clinical breast MRI data for 32 subjects. The results demonstrate that the proposed method achieves a comparable level of performance (AUC of 0.8989± 0.0021 versus AUC of 0.9330± 0.0018). However the advantage of the proposed method over the competing method is that it does not require subjective specification of feature ranges for computing co-occurrence.