PURPOSE: Computerized detection of prostate cancer on T2-weighted MR images. METHODS: The authors combined fractal and multifractal features to perform textural analysis of the images. The fractal dimension was computed using the Variance method; the multifractal spectrum was estimated by an adaptation of a multifractional Brownian motion model. Voxels were labeled as tumor/nontumor via nonlinear supervised classification. Two classification algorithms were tested: Support vector machine (SVM) and AdaBoost. RESULTS: Experiments were performed on images from 17 patients. Ground truth was available from histological images. Detection and classification results (sensitivity, specificity) were (83%, 91%) and (85%, 93%) for SVM and AdaBoost, respectively. CONCLUSIONS: Classification using the authors' model combining fractal and multifractal features was more accurate than classification using classical texture features (such as Haralick, wavelet, and Gabor filters). Moreover, the method was more robust against signal intensity variations. Although the method was only applied to T2 images, it could be extended to multispectral MR.
PURPOSE: Computerized detection of prostate cancer on T2-weighted MR images. METHODS: The authors combined fractal and multifractal features to perform textural analysis of the images. The fractal dimension was computed using the Variance method; the multifractal spectrum was estimated by an adaptation of a multifractional Brownian motion model. Voxels were labeled as tumor/nontumor via nonlinear supervised classification. Two classification algorithms were tested: Support vector machine (SVM) and AdaBoost. RESULTS: Experiments were performed on images from 17 patients. Ground truth was available from histological images. Detection and classification results (sensitivity, specificity) were (83%, 91%) and (85%, 93%) for SVM and AdaBoost, respectively. CONCLUSIONS: Classification using the authors' model combining fractal and multifractal features was more accurate than classification using classical texture features (such as Haralick, wavelet, and Gabor filters). Moreover, the method was more robust against signal intensity variations. Although the method was only applied to T2 images, it could be extended to multispectral MR.
Authors: Radka Stoyanova; Mandeep Takhar; Yohann Tschudi; John C Ford; Gabriel Solórzano; Nicholas Erho; Yoganand Balagurunathan; Sanoj Punnen; Elai Davicioni; Robert J Gillies; Alan Pollack Journal: Transl Cancer Res Date: 2016-08 Impact factor: 1.241
Authors: Satish Viswanath; Robert Toth; Mirabela Rusu; Dan Sperling; Herbert Lepor; Jurgen Futterer; Anant Madabhushi Journal: Proc SPIE Int Soc Opt Eng Date: 2013-03-15