Computer aided diagnosis of breast MR

From UMCN Radiology Research

Introduction

Dynamic contrast enhanced MRI (DCE-MRI) is very sensitive (near 100\%) in detecting malignant lesions in the breast. At the right hormonal cycle breast parenchyma shows very little enhancement and any focal area that stands out is suspect. The differential diagnosis is more difficult as benign and malignant lesions both produce focal enhancement. Several DCE-MRI characteristics have been described to distinguish malignant from benign lesions and increase specificity. Malignant lesions often display a stronger relative enhancement-time curve and the curve often shows a so-called late wash-out sign. These signs need not be present everywhere in the lesion but only in a substantial part of the lesion to be labeled malignant. Other features include the edge sharpness, enhancement pattern, lesion volume, and lesion enhancement structure (scattered enhancement or contiguous area). Pharmacokinetic method

Pharmacokinetic (PK) DCE-MRI may help improve breast cancer differentiation by reducing inter patient and inter MR scanner fluctuations. PK tissue parameters are estimated from DCE-MRI data by fitting a compartment model to the observed data. The compartment model is driven by the local plasma concentration curve which needs to be 'removed' (deconvolved). We will term this deconvolution step: patient calibration. The resulting PK parameters then relate only to local tissue (patho-)physiology. We have researched different patient calibration methods and studied the effect on the diagnostic accuracy. We have developped a computer aided diagnosis (CAD) system for breast MR diagnosis and applied it to a dataset of 97 breast cases.

The area under the curve differed significantly between patient calibration methods. Cluster analysis

The analysis of contrast enhancement in lesions on breast MRI requires the determination of a region of interest (ROI). This process is time-consuming and has only limited reproducibility. We present a method for automated determination of a ROI. Mean shift multidimensional clustering (MS-MDC) was employed to divide each lesion into several spatially contiguous clusters based on multiple enhancement parameters. The ROI of each lesion was then defined as the cluster with the highest posterior probability of malignancy, as determined by an iteratively trained classifier. The area under the receiver operator characteristic curve Az, using these probabilities of malignancy, was then employed to estimate the diagnostic performance of contrast enhancement analysis within these ROIs. The diagnostic performance of MS-MDC was compared against a radiologist's final assessment, who used a reduced standardized 6-point score to indicate the level of suspicion for malignancy. Furthermore, we compared performance against a classifier using histogram analysis (HA), which obtains the feature values for each lesion from the 75% percentile in the histogram of all voxels within the delineated lesions. For MS-MDC, we found Az = 0.82 with a 95% confidence interval (CI) of *-*, significantly higher (p=*) than HA, with Az = 0.71 . For the radiologist's assessment, we found Az = 0.87. There was no statistically significant difference between MS-MDC and the radiologist's performance. We conclude that mean shift multidimensional clustering was capable of automatically determining accurate ROIs in breast MRI lesions.

Area under the curve for clustering based CAD outperformed conventional summary measures.