effects of intensity nonuniformities,nonuniformity correction methods ...
E FFECTS OF I NTENSITY N ONUNIFORMITIES , N ONUNIFORMITY C ORRECTION M ETHODS , AND I NTENSITY N ORMALIZATIONS ON T EXTURE -B ASED C LASSIFICATION OF T2- WEIGHTED A GED C ALF M USCLES Faezeh Fallah1,2 , Nina Schwenzer3 , Fritz Schick2 , Bin Yang1 1 Institute of Signal Processing and System Theory, University of Stuttgart, Germany. 2 Section on Experimental Radiology, University Clinic of Tübingen, Germany. 3 Department of Diagnostic and Interventional Radiology, University Clinic of Tübingen, Germany. M OTIVATION
P URPOSE
VARIABILITY OF T EXTURAL F EATURES WITH N ONUNIFORMITY C ORRECTION AND N ORMALIZATION
Prevalence of aged population in modern societies leads to the rise of health care expenditures due to the significant degradation of life quality of elderly people by aging. Dominance of the muscle tissues in the body (∼40%) and their indispensable roles for respiration, locomotion and body maintenance, motivates development of effective therapeutic means for treatment or reversal of aging-related deterioration of muscle mass and function, called Sarcopenia. This demands the availability of sensitive biomarkers that could differentiate the multi-factorial syndrome of Sarcopenia from other types of pathogenesis with similar phenotypes.
Texture analysis of clinical images is proven to be a promising tool for early detection of pathologies in different tissues. Those images span a wide range of spatial resolution provided by light microscope imagery, pre-clinical high field MRI, and clinical MRI [1, 2]. However, the desired tissue-specific textural features could be confounded with the system-induced tonal features of the imaging apparatus. In this work, we aimed at investigation of the effects of intensity nonuniformities, nonuniformity correction methods, and intensity normalization on texture based classification of T2-weighted MR calf images.
A multi-variate analysis of variance (ANOVA) identified textural features of significant variations with regard to 12 possible combinations of 3 nonuniformity correction and 4 normalization techniques. Following figures show the standard deviations of an exemplary set of run length matrix features computed in the clustered LMM, and the values of two textural features of wavelet and Gabor filters which showed significant variability with regard to the pre-processing operations.
I MAGE A CQUISITION AND P RE - PROCESSING
R ESULTS OF P RE - PROCESSING OF MR I MAGES
MR images were acquired from the calves of 21 asymptomatic volunteers using a T2-weighted turbo spin echo sequence in axial orientation. The used MR scanner was a 3 Tesla MAGNETOM Trio (Siemens Healthineers, Erlangen, Germany). Volunteers’ ages ranged between 19 and 75 years. Other imaging parameters were TR = 2500 ms, TE = 66 ms, slice thickness = 3 mm, and in-plane resolution: 0.42×0.42 mm2 . To correct for intensity nonuniformities, following techniques were applied:
F EATURE D IMENSION R EDUCTION AND C LASSIFICATION E RRORS Three criteria of Probability Of classification Error and Average intra-class Correlation Coefficient (POE+ACC), Fisher (F) metric, and Mutual Information (MI) were used to reduce feature dimension and to classify T2-weighted calf images by a supervised classification method. In the following tables, classification errors resulted from each of those criteria are shown for different combinations of nonuniformity correction and normalization techniques.
• Bias Correction by Fitting samples of Adipose Tissue (BC-FAT) [3] • Bias Correction by fitting samples of Muscle tissue (BCM) [4] • Local Entropy Minimization with a bicubic Spline model (LEMS) [5] To normalize intensities, methods of S1 (no normalization), S2 (multiplicative transformation with regard to maximum intensities), S3 (multiplicative transformation with regard to average intensities), and S4 (restricting intensities to the range between µ − 3σ and µ + 3σ, where µ and σ are the mean and standard deviation of intensities) were used [6]. To compute textural features for lean muscle mass (LMM) and inter-muscular adipose tissues (IMAT) separately, tissues were automatically clustered into these two clusters by using a Gaussian mixture model of intensity histograms and expectation maximization of Gaussian parameters. The optimum number of Gaussian components were identified by an incremental k-means algorithm [7].
R EFERENCES [1] [2] [3] [4] [5] [6] [7]
Lerski et al. EPJ Nonlinear Biomedical Physics, 2015;3:2. De Certaines et al. EPJ Nonlinear Biomedical Physics, 2015;3:3. Würslin et al., JMRI 2011;34:716–726. Fallah et al., ISMRM 2014, p 1176. Salvado et al., IEEE Trans Med Imaging 2006;25(5):539–52. Collewet et al., Magn. Reson. Imag. 2004;22:81–91. Lee et al., IJIT 2006;12(7):13–21.
F UTURE D EVELOPMENTS
D ISCUSSION AND C ONCLUSION In the present study, we performed an automatic clustering of LMM and IMAT, over T2-weighted calf MR images of asymptomatic volunteers with wide range of age variability, by a Gaussian mixture modeling of intensity histograms and expectation maximization in order to separately compute textural features of those segmented tissues. Three methods for correction of intensity nonuniformities were combined with 4 normalization techniques giving 12 possible ways for pre-processing of MR images. A multi-variate ANOVA were used to analyze the sensitivity of textural features to those pre/processing methods and to identity features of significant
variations. Feature dimension reduction and supervised classification of aged calf images were done according to different metrics for each of those 12 data sets. It was found that features of most significant variability were among most discriminant features. Additionally, depending on the criterion used for feature dimension reduction and classification, a certain combination of nonuniformity correction and normalization led to the minimum classification error. This proves high impacts of the pre-processing steps and the classification criterion on texture analysis of MR images.
MR images of multiple contrasts from a large cohort of volunteers with different characteristics enable more advanced techniques of feature dimension reduction, classification, and feature fusion for a better discrimination of aging-induced muscle alterations from other muscular pathologies.
Contact Information: Web http://www.iss.uni-stuttgart.de/institut/mitarbeiter/fallah/ Email [email protected] Phone +49 711 68567360
P URPOSE
VARIABILITY OF T EXTURAL F EATURES WITH N ONUNIFORMITY C ORRECTION AND N ORMALIZATION
Prevalence of aged population in modern societies leads to the rise of health care expenditures due to the significant degradation of life quality of elderly people by aging. Dominance of the muscle tissues in the body (∼40%) and their indispensable roles for respiration, locomotion and body maintenance, motivates development of effective therapeutic means for treatment or reversal of aging-related deterioration of muscle mass and function, called Sarcopenia. This demands the availability of sensitive biomarkers that could differentiate the multi-factorial syndrome of Sarcopenia from other types of pathogenesis with similar phenotypes.
Texture analysis of clinical images is proven to be a promising tool for early detection of pathologies in different tissues. Those images span a wide range of spatial resolution provided by light microscope imagery, pre-clinical high field MRI, and clinical MRI [1, 2]. However, the desired tissue-specific textural features could be confounded with the system-induced tonal features of the imaging apparatus. In this work, we aimed at investigation of the effects of intensity nonuniformities, nonuniformity correction methods, and intensity normalization on texture based classification of T2-weighted MR calf images.
A multi-variate analysis of variance (ANOVA) identified textural features of significant variations with regard to 12 possible combinations of 3 nonuniformity correction and 4 normalization techniques. Following figures show the standard deviations of an exemplary set of run length matrix features computed in the clustered LMM, and the values of two textural features of wavelet and Gabor filters which showed significant variability with regard to the pre-processing operations.
I MAGE A CQUISITION AND P RE - PROCESSING
R ESULTS OF P RE - PROCESSING OF MR I MAGES
MR images were acquired from the calves of 21 asymptomatic volunteers using a T2-weighted turbo spin echo sequence in axial orientation. The used MR scanner was a 3 Tesla MAGNETOM Trio (Siemens Healthineers, Erlangen, Germany). Volunteers’ ages ranged between 19 and 75 years. Other imaging parameters were TR = 2500 ms, TE = 66 ms, slice thickness = 3 mm, and in-plane resolution: 0.42×0.42 mm2 . To correct for intensity nonuniformities, following techniques were applied:
F EATURE D IMENSION R EDUCTION AND C LASSIFICATION E RRORS Three criteria of Probability Of classification Error and Average intra-class Correlation Coefficient (POE+ACC), Fisher (F) metric, and Mutual Information (MI) were used to reduce feature dimension and to classify T2-weighted calf images by a supervised classification method. In the following tables, classification errors resulted from each of those criteria are shown for different combinations of nonuniformity correction and normalization techniques.
• Bias Correction by Fitting samples of Adipose Tissue (BC-FAT) [3] • Bias Correction by fitting samples of Muscle tissue (BCM) [4] • Local Entropy Minimization with a bicubic Spline model (LEMS) [5] To normalize intensities, methods of S1 (no normalization), S2 (multiplicative transformation with regard to maximum intensities), S3 (multiplicative transformation with regard to average intensities), and S4 (restricting intensities to the range between µ − 3σ and µ + 3σ, where µ and σ are the mean and standard deviation of intensities) were used [6]. To compute textural features for lean muscle mass (LMM) and inter-muscular adipose tissues (IMAT) separately, tissues were automatically clustered into these two clusters by using a Gaussian mixture model of intensity histograms and expectation maximization of Gaussian parameters. The optimum number of Gaussian components were identified by an incremental k-means algorithm [7].
R EFERENCES [1] [2] [3] [4] [5] [6] [7]
Lerski et al. EPJ Nonlinear Biomedical Physics, 2015;3:2. De Certaines et al. EPJ Nonlinear Biomedical Physics, 2015;3:3. Würslin et al., JMRI 2011;34:716–726. Fallah et al., ISMRM 2014, p 1176. Salvado et al., IEEE Trans Med Imaging 2006;25(5):539–52. Collewet et al., Magn. Reson. Imag. 2004;22:81–91. Lee et al., IJIT 2006;12(7):13–21.
F UTURE D EVELOPMENTS
D ISCUSSION AND C ONCLUSION In the present study, we performed an automatic clustering of LMM and IMAT, over T2-weighted calf MR images of asymptomatic volunteers with wide range of age variability, by a Gaussian mixture modeling of intensity histograms and expectation maximization in order to separately compute textural features of those segmented tissues. Three methods for correction of intensity nonuniformities were combined with 4 normalization techniques giving 12 possible ways for pre-processing of MR images. A multi-variate ANOVA were used to analyze the sensitivity of textural features to those pre/processing methods and to identity features of significant
variations. Feature dimension reduction and supervised classification of aged calf images were done according to different metrics for each of those 12 data sets. It was found that features of most significant variability were among most discriminant features. Additionally, depending on the criterion used for feature dimension reduction and classification, a certain combination of nonuniformity correction and normalization led to the minimum classification error. This proves high impacts of the pre-processing steps and the classification criterion on texture analysis of MR images.
MR images of multiple contrasts from a large cohort of volunteers with different characteristics enable more advanced techniques of feature dimension reduction, classification, and feature fusion for a better discrimination of aging-induced muscle alterations from other muscular pathologies.
Contact Information: Web http://www.iss.uni-stuttgart.de/institut/mitarbeiter/fallah/ Email [email protected] Phone +49 711 68567360
