A ranklet‐based image representation for mass classification in digital mammograms

Abstract

Regions of interest ͑ROIs͒ found on breast radiographic images are classified as either tumoral mass or normal tissue by means of a support vector machine classifier. Classification features are the coefficients resulting from the specific image representation used to encode each ROI. Pixel and wavelet image representations have already been discussed in one of our previous works. To investigate the possibility of improving classification performances, a novel nonparametric, orientation-selective, and multiresolution image representation is developed and evaluated, namely a ranklet image representation. A dataset consisting of 1000 ROIs representing biopsy-proven tumoral masses ͑either benign or malignant͒ and 5000 ROIs representing normal breast tissue is used. ROIs are extracted from the digital database for screening mammography collected by the University of South Florida. Classification performances are evaluated using the area A z under the receiver operating characteristic curve. By achieving A z values of 0.978± 0.003 and 90% sensitivity with a false positive fraction value of 4.5%, experiments demonstrate classification results higher than those reached by the previous image representations. In particular, the improvement on the A z value over that achieved by the wavelet image representation is statistically relevant with the two-tailed p value Ͻ0.0001. Besides, owing to the tolerance that the ranklet image representation reveals to variations in the ROIs' gray-level intensity histogram, this approach discloses to be robust also when tested on radiographic images having gray-level intensity histogram remarkably different from that used for training.

Texture classification using invariant ranklet features

“…Each image region is submitted to the ranklet transform and decomposed into n R = 7 ranklet resolutions (i.e., {4, 6, 8, 10, 12, 14, 26}) and n O = 3 orientations (i.e., vertical, horizontal, and diagonal). Similarly to what discussed in one of our previous works (Masotti, 2006b), this choice is as arbitrary as reasonable, since it spans over a large range of resolutions, from fine ones (encoding closeview texture details) to coarse ones (encoding broad-view texture details). In particular, preliminary tests confirmed the validity of the selected ranklet resolutions.…”

“…First, given an image I, the ranklet transform is applied (Smeraldi, 2002). As discussed in some of our recent works (Masotti, 2006a;Campanini et al, 2006;Masotti, 2006b), in fact, being calculated from the relative rank of pixels rather than from their gray-scale value (i.e., non-parametric property of the ranklet transform), this transform allows to produce a grayscale invariant image representation; more specifically, as it will be shown in the following, gray-scale invariance is intended as to linear/non-linear monotonic gray-scale transformations of the original image I, e.g., brigthness variation, contrast enhancement, gamma correction, histogram equalization (Gonzalez and Woods, 1992). Also, being calculated as a multi-resolution and orientation-selective analysis (i.e., multi-resolution and orientation-selective properties of the ranklet transform), this transform allows to recognize analogous characteristics at different resolutions and orientations of the image as well.…”

Exploring Ranklets Performances in Mammographic Mass Classification using Recursive Feature Elimination

“…The total number of images used amounts to 6000 and is partitioned in 1000 images representing the mass class and 5000 images representing the non-mass class. Notice that the images used in this paper are exactly those used to evaluate the classification performances in our previous works [2,3]. In Fig.…”

“…Finally, the 1428 ranklet-based features are obtained resizing the original 64 × 64 pixels images to 16 × 16 pixels, applying a multi-resolution ranklet transform and thus taking the ranklet coefficients: in particular, multi-resolution is achieved by stretching the Haar wavelet supports to dimensions 16 × 16, 8 × 8, 4 × 4 and 2 × 2 pixels. For more details see [2,3]. The performances are compared using Receiver Operating Characteristic (ROC) curves generated by moving the hyperplane of the SVM solution.…”

“…Masses are thickenings of the breast tissue with size ranging from 3 mm to 30 mm often associated with the presence of a breast tumor. Similarly to what has been done in a couple of our previous works [2,3], given a dataset composed of images representing both the mass class and the non-mass class, the ranklet transform is taken on each image, then the transformed images are classified by means of a trained Support Vector Machine (SVM). The novelty in this paper is that, in order to reduce the great amount of ranklet coefficients arising from the ranklet transform of each image, a feature reduction technique, known as Recursive Feature Elimination (RFE), is applied.…”