Improved lesion detection in MR mammography: three-dimensional segmentation, moving voxel sampling, and normalized...

The objective of this work was to develop a quantitative method for improving lesion detection in dynamic contrast-enhanced magnetic resonance mammography (DCEMRM). For this purpose, we segmented and analyzed suspicious regions according to their contrast enhancement dynamics, generated a normalized maximum intensity-time ratio (nMITR) projection, and explored it to extract important features, to improve accuracy and reproducibility of detection.

A novel automated method is introduced to segment and analyze lesions in three dimensions. It consists of four consecutive stages: volume of interest selection, nMITR projection generation using a voxel sampling method based on a moving 3 x 3 mask, three-dimensional lesion segmentation, and feature extraction. The nMITR projection of the detected lesion is used to extract six features: mean, maximum, standard deviation, kurtosis, skewness, and entropy, and their diagnostic significance is studied in detail. High-resolution MR images of 52 breast masses from 46 women are analyzed using the technique developed.

Entropy, standard deviation, and the maximum and mean value features were found to have high significance (P < 0.001) and diagnostic accuracy (0.86-0.97). The kurtosis and skewness were not significant. Automated analysis of DCEMRM using nMITR was shown to be feasible.

The lesion detection method described is efficient and leads to improved, accurate, reproducible diagnoses. It is reliable in terms of observer variability and may allow for a better standardization of clinical evaluations. The findings demonstrate the usefulness of nMITR based features; nMITR-entropy shows the best performance for quantitative diagnosis.