| Περίληψη: | Magnetic resonance mammography (MRM) is a promising technique, since it provides high resolution breast imaging with no use of ionising radiation and with inherently good soft tissue discrimination. The addition of dynamic contrast enhancement kinetics of the breast upgraded the method to a great extend, due to highly differentiated malignant vs. benign lesion hemodynamics resulting from the angiogenetic properties of cancerous cells.
Straightforward pharmacokinetic analysis, such as the 3TP algorithm, has been implemented in commercially available CAD systems. Quantitative parameters can be extracted that directly correspond to different aspects of the underlying pathology and can be compared to biopsy results. However, there is a general understanding that straightforward pharmacokinetic analysis (3TP model) requires a very demanding imaging protocol in order to be able to measure such parameters accurately. Fitting the experimental data of the dynamic series to simple mathematical models extracting quantitative features provides a means to evaluate and to shrink the big amount of data of the study to one set of images, in order make the diagnostic process faster and more robust. That could facilitate the clinical routine.
The dynamic series of the MRM examinations of the 55 patients were analyzed in this study. Radiologists specialized in MRM have identified and characterized all suspicious lesions according to BIRADS lexicon. Dynamic data were fitted pixel-wise to a simple bilinear model to extract washout, time to peak and washin parameters. Subsequently, those parameters were mapped to Hue, Saturation and Value, respectively, of an HSV color model, which was utilized for characterizing the lesions. Also, Hue heterogeneity was qualitatively assessed for the characterization of lesions. In addition, observers evaluated the haemodynamic properties of the lesions with the conventional hand-drawn ROI based technique (Kuhl system).
The results of the two methods were then compared to the histological ground truth to derive their classification performance. Classification performance for the proposed and the conventional one was Az=0.880.05 and Az=0.860.05, respectively, by means of ROC analysis. Results indicate no statistically different performance between the two methods, with the proposed one offering time savings and reproducibility.
|
|---|
