Classification of Dynamic Breast MR Image Data - Biomedical Signal Analysis: Contemporary Methods and Applications - page 285

Table 10.1
Comparison of different data-driven segmentation methods of dynamic
contrast-enhanced breast MRI time series. The differentiation between benign and
malignant lesions is based on the method described in [145]. m is a malignant lesion
and b a benign lesion.

Method I ∗

Data set Method I Method III Lesion Description

#1 III III III m Scirrhous carcinoma

#2 II II III m Tubulo-lobular carcinoma

#3 Ib Ib Ib b Fibroadenoma

#4 Ib Ib III m Ductal carcinoma in situ

#5 Ia Ia Ia b Fibrous mastopathy

#6 III III III m Papilloma

#7 II II II m Ductal carcinoma in situ

#8 Ib Ib Ib b Inflammatory granuloma

#9 Ib Ib Ib b Scar, no relapse

#10 Ib Ib II m Ductal carcinoma in situ

#11 II II III m Invasive, ductal carcinoma

#12 Ib Ib Ib b Fibroadenoma

#13 III III III m Medullary carcinoma

characteristic of benign lesions. This fact is visualized in figure 10.11. The
resulting mismatch between these two segmentation methods shows the
main advantage of segmentation method III: based on a differentiated
examination of tissue changes, we obtain an increase in sensitivity of
breast MRI with respect to malignant lesions.
The examined data sets show that the relevance of the minimal free
energy vector quantization neural network for MRI breast examination
lies in the potential to increase the diagnostic accuracy for MRI mam-
mography by improving the sensitivity without reduction of specificity.
In order to document this improvement induced by segmentation method
III, the results are included of all three segmentation methods on all the
“critical” data sets (i.e., those where such a mismatch between segmen-
tation methods I and III could be observed: data sets #2, 4, 10, and 11),
see figures 10.7-10.22.
In this chapter, three different segmentation methods have been
presented for the evaluation of signal-intensity time-courses for the
differential diagnosis of enhancing lesions in breast MRI. Starting from
the conventional methodology, the concepts of threshold segmentation
and cluster analysis were introduced and in the last step those two
concepts were combined.
The introduction of new techniques was motivated by the conceptual