Texture Based Computer Aided Malignant Lesion Segmentation of MR Mammography Images Compared With Majority of Manual Segmentations

Results After training the neural network, all voxels from all seven datasets were classified and both color-coded images and statistical results were obtained. Figure 1 shows the color-coded images comparing the classifier output with the gold standard majority segmentations. From the images we notice that the majority of misclassifications (FP,FN) occur near the border of the lesion where partial volume effects may influence results. Figure 2 shows the ROC analysis based on the TPF and FPF calculated from all voxels in all seven studies. In addition to classifier performance the plot also indicates how each radiologist performed at each of the two times (two weeks apart) compared with the gold standard majority segmentations. These results indicate inter- and intra-observer variations relative to the neural network performance. A paired t-test at significance level of 0.05 was performed using the TPF and FPF values obtained from each of the seven studies. The classifier performance relative to the gold standard was tested against each of the manual segmentations relative to the gold standard. Forty-two total paired samples were obtained (six segmentations per subject times seven subjects). The t-test results show that at thresholds between 30% to 60%, the classifier performs statistically equal to or better on average than the manual segmentations relative to the gold standard majority segmentations. At a threshold of 50%, the classifier performs statistically better on average than the manual segmentations relative to the gold standard majority segmentations. Conclusion Texture analysis along with a neural network based classifier can be used to detect malignant lesions in DCE-MRI datasets. The results of this study show that the classifier can be trained to emulate the segmentation performance of the majority of a group of radiologists. Further it is shown that once trained the classifier can on average perform on par or better in the statistical (TPF,FPF) sense than manual segmentations. This is because the classifier gives consistent computational results whereas inter- and intra-observer variations occur within the manual segmentation process. When used as a tool, the classifier gives the radiologists a first guess approximation of malignant lesion size and location. From there the radiologist may adjust the segmentation to his or her preference.