Abstract:
In this thesis, algorithms, methods and techniques for dynamic contrast-enhanced magnetic resonance mammography (DCE-MRM) have been investigated to maximize sensitivity, specificity and reproducibility of breast cancer diagnoses. A novel lesion localization method that uses cellular neural networks (CNNs) was developed. The breast region was segmented from pre-contrast images using four specifically designed CNNs. A 3D normalized maximum intensity-time ratio (nMITR) map of the segmented breast was generated using a moving mask of 3×3 voxels on the dynamic images. This map was converted into a binary form and processed with a fuzzy CNN consisting of three layers of 11×11 cells to segment out lesions from the surrounding tissues and to filter-out deceptive enhancements. A set of decision rules based on volume and 3D eccentricity of the suspicious regions were applied to minimize false-positive detections. The system was tested on a dataset consisting of 7020 MR mammograms in 1170 slices from 39 patients with 37 malignant and 39 benign mass lesions and was found to perform well with falsepositive detections of 0.34/lesion, 0.10/slice and 0.67/case at a maximum detection sensitivity of 99%. Enhancement and morphological descriptors of breast lesions derived from 3D nMITR maps were also studied for malignancy detection. The mean, the maximum value, the standard deviation and the entropy were the enhancement features found to have high significance (P< 0.001) and diagnostic accuracy (0.86-0.97). nMITR-entropy had the best performance. Among the morphological descriptors studied, 3D convexity, complexity and extent were found to have higher diagnostic accuracies (ranging between 0.70-0.81) and better performance than their 2D versions. Contact surface area ratio was found to be the most significant and accurate descriptor (75% sensitivity, 88% specificity, 89% PPV and 74% NPV). The results demonstrate that nMITR maps inherently suppress enhancements due to normal parenchyma and blood vessels that surround lesions and have natural tolerance to small field homogeneities and thus are very effective for lesion localization and malignancy detection.
