Meningiomas are the most common primary intracranial tumors in dogs.1 Although histologic grading of these tumors is important for therapeutic and prognostic purposes, no specific MRI features are reported to be correlated with histologic subtype or grade.1 Texture analysis is the mathematical description of the relationship between pixel intensities (eg, gradient values, homogeneity, and random distribution) within a given ROI of an image. This technique is used to increase the amount of information that can be obtained from diagnostic images, thereby overcoming the limited ability of the human eye to distinguish among various shades of gray.2 Possible applications of TA to MRI sequences have been widely investigated in human medicine,2 and several reports are available on this topic. The main fields of research are focused on the staging of intracranial tumors,3–5 evaluation of brain degenerative diseases,6–8 liver fibrosis,9,10 thyroid tumors,11 and response of non-Hodgkin lymphoma to chemotherapy.12
A branch of quantitative analysis of medical images, TA has been seldom proposed for use in veterinary medicine, and existing proposals mainly include characterization of normal renal and hepatic echogenicity.13,14 The correlation between renal lesions and renal echogenicity15 and the relationship between hepatic degenerative changes and hepatic echogenicity16 have been investigated. Moreover, only a few reports17–20 exist in the veterinary medical literature of clinical applications of TA to medical images.
The purpose of the study reported here was to develop a fast and reliable procedure for performing TA on MRI sequences of the brain and spinal cord in dogs and to determine whether TA could be used to predict histologic grade of meningiomas. To the authors’ knowledge, this would represent the first time such analysis was explored in veterinary medicine.
Materials and Methods
Sample
Fifty-eight sets of MRI scans of the brain and spinal cord of canine patients of the PRVH (n = 27) and DWR (31) with a final histologic diagnosis of meningioma were retrospectively selected. Scans from dogs evaluated at the PRVH were used to develop the statistical model (training set) for prediction of histologic grade of meningioma, whereas scans from dogs evaluated at the DWR were used to test the results of this statistical model test set (Figure 1).
Flowchart of procedures used to determine whether histologic grade of meningiomas in dogs could be predicted through TA of MRI scans of the brain and spinal cord. The MRI scans of patients at one hospital were used to develop a statistical model (training set; A), and MRI scans of patients at another hospital were used to test the model (test set; B). The B11 is a software application.
Citation: American Journal of Veterinary Research 78, 10; 10.2460/ajvr.78.10.1156
Image acquisition protocols
The MRI scans from DWR had been acquired with a 0.4-T open permanent magnet,b and those from PRVH with a 0.22-T open permanent magnet.c Different protocols were used. However, all MRI examinations included 3 sequences: a fast spin echo T2-weighted series (repetition time, 13 to 120 milliseconds; echo time, 290 to 7,790 milliseconds; matrix, 512 × 512 pixels) and pre- and postcontrast (gadolinium-based medium) spin echo T1-weighted series (repetition time, 13 to 26 milliseconds; echo time, 462 to 880 milliseconds; matrix, 512 × 512 pixels) in the transverse plane. All images were acquired with 3- to 5-mm slice thickness with a 10% gap, and the signal-to-noise ratio was improved by the use of 2 to 4 averages for each acquisition.
Histologic assessment and tumor classification
Data were collected from the medical records regarding histologic grade of meningioma. All histologic examinations for dogs at DWR had been performed by board-certified pathologists. All histologic examinations for dogs at PRVH had been performed by an experienced pathologist at the University of Perugia, Italy. Classifications had been assigned by use World Health Organization guidelines21 including grade I (benign; histologic variant other than clear cell, chordoid, papillary, and rhabdoid; lacking criteria of atypical and anaplastic meningioma), grade II (atypical; mitotic index, ≥ 4 mitoses/10 hpfs; at least 3 of the following 5 characteristics: sheeting architecture, small cell formation, macronucleoli, hypercellularity, and spontaneous necrosis; brain invasion), and grade III (anaplastic; mitotic index, ≥ 20 mitoses/10 hpfs; frank anaplasia [sarcoma, carcinoma, or melanomalike histologic features]).
TA
Texture analysis of selected MRI scans was performed with dedicated softwarea following the developers’ indications (co-occurrence matrix, 6 bits/pixel; gradient features, 8 bits/pixel; run-length matrix, 4 bits/pixel; wavelet transform, 12 bits/pixel). All possible texture parameters were calculated (histogram, gradient map, co-occurrence matrix, run-length matrix, autoregressive model, and wavelet transform; Appendix 1). The TA outcome was normalized by rescaling of the data to fit 3 SDs above and below the mean intensity level.
The 3 MRI sequences were imported directly in the TA software in DICOM format. All lesions were first identified in postcontrast T1-weighted images. A polygonal ROI was manually placed on the lesion, and texture parameters were then calculated; thereafter, the corresponding precontrast T1-weighted image (in which the lesion was usually barely visible) was imported into a TA program so that the preselected ROI precisely overlapped the previously localized lesion (Figure 2). The same operation was then performed with T2-weighted images. To account for intratumor variability and to increase the amount of data available for the statistical analysis, 3 consecutive images for each lesion (when available) for each MRI sequence were selected for every dog. Images in which the lesion was larger or more easily identifiable were always selected. Texture analysis was not performed on MRI scans in which the lesion was not clearly identifiable. Non–contrast-enhancing regions in postcontrast T1-weighted images were carefully excluded from TA as possible intratumoral necrosis. The TA results for the 3 MRI sequences were analyzed separately.
Transverse MRI views of the head of an 8-year-old German Shepherd Dog with a grade III papillary meningioma demonstrating the workflow of ROI selection for TA. A—A postcontrast T1-weighted image was loaded into the TA software. B—A polygonal ROI following the leading edge of the lesion was selected, and TA was performed. C—The corresponding precontrast T1-weighted image was thereafter loaded. A perfect match between the different image sequences was achieved, thereby enabling direct application of an ROI selected in postcontrast T1-weighted images to the other sequences. D—The same procedure as in panel C was performed for T2-weighted images. Bar = 1 cm.
Citation: American Journal of Veterinary Research 78, 10; 10.2460/ajvr.78.10.1156
Because the TA procedure involves calculation of a large number of image features (279 with existing program settings), the TA software enables several feature-reduction methods to identify image characteristics that yield the most accurate classification of samples (in this situation, benign vs atypical-anaplastic meningiomas). The combination of the Fisher coefficient, classification error probability combined with mean correlation coefficients, and mutual information algorithms, selecting a total of 30 texture features, was used (Appendix 2). A complete description of the texture features is reported elsewhere.22–26
Training set data analysis
The texture features selected by means of the reduction algorithms for MRI scans of the PRVH were recorded and imported into the B11 application (embedded in the TA software) for data analysis and classification. The most discriminating texture features (Appendix 2) were automatically selected by the TA software from the training set by use of the embedded selection algorithms. To test the classification efficiency of the TA features, an LDA was then performed with use of the software B11 application, and the misclassification rate was recorded. The MDF values resulting from the LDA model were then exported to commercially available statistical softwared to generate ROC curves.
Test set data analysis
The most discriminating texture features automatically selected for the training set for each of the 3 MRI sequences were manually selected from the TA results for the test set (scans from DWR) and imported in the B11 program. The LDA was performed by use of these data, and MDF values were calculated. Cutoff MDF values for each MRI sequence calculated by use of the training set were used in classifying the images in the test set. The misclassification rate was manually calculated.
Results
Pathologist-assigned histologic scores for meningiomas in the training and test sets, respectively, included grade I (benign; n = 8 and 20), grade II (atypical; 15 and 6), and grade III (anaplastic; 4 and 5). Six tumors (1 grade I and 5 grade III meningiomas) were cystic with only a small amount of tissue available for analysis and were therefore excluded from the TA. Because only 4 grade III tumors were consequently available for TA, these tumors were incorporated with grade II tumors for statistical analysis, resulting in a binary histologic classification system (benign vs atypical-anaplastic).
Histotypes of the meningiomas in the training and test sets, respectively, included papillary (n = 6 and 5), transitional (2 and 8), fibroblastic (3 and 3), atypical (4 and 2), chordoid (3 and 0), meningothelial (2 and 4), anaplastic (2 and 0), psammomatous (0 and 2), and other (5 and 7). Other types included 1 each of biphasic, cystic, myxoid, infiltrative, lipomatous, malignant, microcystic, osteoid, secretory, syncytial, vacuolar, and vascular.
Scans for 4 dogs (1 from the training set and 3 from the test set) were excluded from the TA because the lesions were too small or poorly visible in postcontrast T1-weighted images. The final training set comprised 72 images (20 images classified as benign [n = 7 tumors] and 52 classified as typicalanaplastic [18 tumors]), whereas the test set comprised 62 images (41 images classified as benign [16 tumors] and 21 as atypical-anaplastic [7 tumors]).
Linear discriminant analysis of the training set samples of precontrast T1-weighted images yielded a misclassification rate of 8.4% for TA in discrimination between benign and atypical-anaplastic meningiomas (Figure 3). Area under the ROC curve for the TA MDF in predicting histologic classification was 0.986 (95% CI, 0.895 to 0.998). When a cutoff MDF value of < 0.0057 was applied, sensitivity and specificity of TA for predicting histologic classification were 97.4% and 95.0%, respectively, and the positive likelihood ratio was 19.5. The MDF of the test set was calculated through LDA; setting a cutoff MDF value of < 0.0057 yielded a misclassification rate of 3.2% (correct classification rate of 96.8%).
Output of a TA analysis program (B11) for LDA involving precontrast (A) or postcontrast (C) T1-weighted or T2-weighted (E) images in the training set of canine brain and spinal cord MRI scans and associated ROC curves (B, D, and F, respectively) for the TA MDF in the prediction of histologic grade of meningioma (benign vs atypical-aplastic). A—A misclassification rate of 8.4% was registered (overlap between the red and green numbered series). B—Area under the ROC curve was 0.986 (95% CI, 0.895 to 0.998). When a cutoff MDF value of < 0.0057 was used, sensitivity of TA for distinguishing benign from atypical-anaplastic tumors was 97.4%, specificity was 95.0%, and the positive likelihood ratio was 19.49. C—A misclassification rate of 13.9% was registered. D—Area under the ROC curve was 0.980 (95% CI, 0.915 to 0.999). When a cutoff MDF value of < 0.001 was used, sensitivity was 94.7%, specificity was 94.3%, and the positive likelihood ratio was 16.7. E—A misclassification rate of 15.9% was registered. F—Area under the ROC curve was 0.966 (95% CI, 0.915 to 0.999). When a cutoff MDF value of < 0.004 was used, sensitivity was 100%, specificity was 86.0%, and the positive likelihood ratio was 7.1. Red numbers in panels A, C, and E represent meningiomas histologically classified as benign, whereas green numbers represent meningiomas histologically classified as atypical or anaplastic.
Citation: American Journal of Veterinary Research 78, 10; 10.2460/ajvr.78.10.1156
Linear discriminant analysis of postcontrast T1-weighted image samples provided a misclassification rate of 13.9% (Figure 3). Area under the ROC curve for the TA MDF was 0.980 (95% CI, 0.915 to 0.999). When a cutoff MDF value of < 0.001 was applied, sensitivity and specificity were 94.7% and 94.3%, respectively, and the positive likelihood ratio was 16.7. The MDF of the test set was calculated through LDA; setting a cutoff MDF value of < 0.001 yielded a misclassification rate of 8.0% (correct classification rate of 92.0%).
Linear discriminant analysis of the T2-weighted image samples yielded a misclassification rate of 15.9% (Figure 3). Area under the ROC curve for the TA MDF was 0.966 (95% CI, 0.915 to 0.999). When a cutoff MDF value of < 0.004 was used, sensitivity and specificity were 100% and 86.0%, respectively, and the positive likelihood ratio was 7.1. The MDF of the test set was calculated through LDA; setting a cutoff MDF value of < 0.004 yielded a misclassification rate of 21.1% (correct classification rate of 78.9%).
Discussion
The TA procedure evaluated in the present study enabled fast and consistent selection of lesions from the different MRI sequences of the brain and spinal cord in dogs with meningioma. In veterinary medicine, MRI examinations are usually performed with animals anesthetized, and a perfect match between the different image sequences can be consistently achieved, thereby enabling direct application of an ROI selected in postcontrast T1-weighted images to the other sequences. An ROI selection procedure similar to that used in the present study has been used to evaluate response to chemotherapy in humans with non-Hodgkin lymphoma.27 Moreover, the same texture features used in the present study could also be calculated by means of other commercially available software. The TA software used in the study was chosen because it was specifically developed for analysis of MRI scans.28,29
The training and test sets of MRI scans had different distributions of histologic categories of tumors, most likely owing to the small sample size, with the training set (PRVH) having a higher number of atypical-anaplastic tumors and the test set (DWR) having a higher number of benign tumors. Furthermore, the 2 institutions from which scans were obtained had different MRI scanners (0.22 T at PRVH and 0.4 T at DWR). However, we believe that such differences strengthened the results of the study given that, despite dissimilar distributions of tumor types and different types of MRI scanner used, high classification efficiency was obtained (ie, 96.8% of T1-weighted images in the test set were correctly classified). Interestingly, even in human medicine, different MRI scanners provide similar TA results.11
A possible limitation of the study reported here was that the relatively low number of dogs in the test set and relatively high number of texture parameters used to perform LDA may have resulted in an overfitting of the obtained model. Studies involving a larger number of MRI scan sets are required to test the obtained classification efficiency.11 Given that results achieved with the training set and test set were similar, overfitting did not appear to have affected our model.
Another limitation of the present study was indirectly indicated by the fairly high number of MRI scan sets discarded from the analysis (representing 10 of 58 [17%] dogs). Six scan sets were discarded because of the cystic nature of the lesions, which resulted in only a minimal amount of tissue available for the TA, whereas 4 additional scan sets were discarded because the lesion was not clearly visible in postcontrast T1-weighted images. The aforementioned limitations suggest that the procedures reported here could be applied only to noncystic and well-defined lesions.
Lastly, because of the low prevalence of grade III meningiomas in the study sample, atypical and anaplastic tumors were grouped together as a single category. Additional studies including a larger number of cases are required to determine the usefulness of TA for discriminating between atypical and anaplastic meningiomas.
Most meningiomas in dogs appear as heterogeneous in precontrast T1-weighted images,1 and, not surprisingly, the best classification results for the test set in the present study were obtained with this sequence. Texture analysis quantifies tissue heterogeneity30,31 through evaluation of local signal intensity variations (usually imperceptible to the human visual system). The general principle supporting TA use for the staging of neoplasms is the positive correlation between the degree of malignancy of a neoplasm and the level of destructuration of normal tissue, which results in a greater heterogeneity in the MRI scans.2,31 Intravenous administration of contrast medium enhances visibility of the lesions but makes the target tissue more homogeneous so that local variations in signal intensity are reduced, thereby providing poorer TA performance.
Good classification results are reportedly achieved in human medicine for the combination of TA and dynamic contrast-enhanced MRI in the differentiation between high- and low-grade gliomas.32 Canine meningiomas mainly appear as homogeneous in T2-weighted images,1 therefore providing poor information about the actual structure of the tissue. This could at least partially explain the less accurate TA classification results obtained with T2-weighted versus T1-weighted images in the present study. Intratumoral mineralizations have been identified in dogs with meningioma,21 but we observed no such mineralizations. Given that mineralizations produce a high signal heterogeneity despite lesion malignancy, it is the authors’ opinion that mineralized areas should never be included in a TA. Likewise, non–contrast-enhancing areas were carefully excluded from the analysis as possible expressions of intratumoral necrosis. Nevertheless, we acknowledge that contrast enhancement might occur also within necrotic areas.33
Treatment options for dogs with meningioma basically consist of surgery, radiation therapy, or a combination of both. In human medicine, treatment of meningiomas is based on histologic grade of tumors because atypical and malignant lesions have a higher recurrence rate than benign variants.1 To date, no definitive data are available on the treatment of choice for tumors by type or location, despite a recent study34 having revealed no differences in the survival times of meningiomaaffected dogs treated with 3-D conformal radiation therapy alone and in combination with surgery. Nevertheless, a possible relationship has been proposed between histologic grade of meningiomas and response to treatment in dogs.1 The possibility to grade meningiomas directly from MRI scans could assist clinicians when choosing treatments for their patients or researchers when assessing the relationship between treatment options and tumor malignancy.
Findings of the study reported here did not suggest that TA can be immediately clinically applied as a diagnostic test. In fact, data from a preexisting database of histologically confirmed meningiomas and corresponding MRI scans would be required to build upon the predictive model developed here. Future studies involving larger databases could enable the development of dedicated software for providing an immediate, simple, precise, and accurate diagnosis in dogs with meningioma.
Acknowledgments
This report is part of a pilot study project funded by 2 research grants from the University of Padova, Italy (No. CPDA124900/2012 and Junior Research Grant 2015), and by a research grant from Veneto Regional Council, Italy (No. FSE 2105/201/27/1148/2013).
ABBREVIATIONS
| CI | Confidence interval |
| DWR | Dick White Referrals |
| LDA | Linear discriminant analysis |
| MDF | Most discriminant factor |
| PRVH | Portoni Rossi Veterinary Hospital |
| ROC | Receiver operating characteristic |
| ROI | Region of interest |
| TA | Texture analysis |
Footnotes
MaZda, version 4.6, Technical University of Lodz, Institute of Electronics, Lodz, Poland.
Hitachi Aperto, Hitachi Medical Corp, Tokyo, Japan.
MrV, Paramed Medical Systems, Genova, Italy.
MedCalc for Windows, version 12.5, MedCalc Software, Ostend, Belgium.
References
1. Sturges BK, Dickinson PJ, Bollen AW, et al. Magnetic resonance imaging and histological classification of intracranial meningiomas in 112 dogs. J Vet Intern Med 2008; 22: 586–595.
2. Castellano G, Bonilha L, Li LM, et al. Texture analysis of medical images. Clin Radiol 2004; 59: 1061–1069.
3. Orphanidou-Vlachou E, Vlachos N, Davies NP, et al. Texture analysis of T1- and T1-weighted MR images and use of probabilistic neural network to discriminate posterior fossa tumours in children. NMR Biomed 2014; 27: 632–639.
4. Mahmoud-Ghoneim D, Toussaint G, Constans JM, et al. Three dimensional texture analysis in MR: a preliminary evaluation in gliomas. Magn Reson Imaging 2003; 21: 983–987.
5. Zacharaki EI, Wang S, Chawla S, et al. Classification of brain tumor type and grade using MR texture and shape in a machine learning scheme. Magn Reson Med 2009; 62: 1609–1618.
6. de Albuquerque M, Anjos LGV, Maia Tavares de Andrade H, et al. MR texture analysis reveals deep gray nuclei damage in amyotrophic lateral sclerosis. J Neuroimaging 2016; 26: 201–206.
7. Bonilha L, Kobayashi E, Castellano G, et al. Texture analysis of hippocampal sclerosis. Epilepsia 2003; 44: 1546–1550.
8. Sikiö M, Holli-Helenius KK, Harrison LC, et al. MR image texture in Parkinson's disease: a longitudinal study. Acta Radiol 2015; 56: 97–104.
9. Bahl G, Cruite I, Wolfson T, et al. Noninvasive classification of hepatic fibrosis based on texture parameters from double contrast-enhanced magnetic resonance images. J Magn Reson Imaging 2012; 36: 1154–1161.
10. Yeom SK, Lee CH, Cha SH, et al. Prediction of liver cirrhosis, using diagnostic imaging tools. World J Hepatol 2015; 7: 2069–2079.
11. Brown AM, Nagala S, McLean MA, et al. Multi-institutional validation of a novel textural analysis tool for preoperative stratification of suspected thyroid tumors on diffusion-weighted MR. Magn Reson Med 2016; 75: 1708–1716.
12. Harrison L, Dastidar P, Eskola H, et al. Texture analysis on MR images of non-Hodgkin lymphoma. Comput Biol Med 2008; 38: 519–524.
13. Drost WT, Henry GA, Meinkoth JH, et al. Quantification of hepatic and renal cortical echogenicity in clinically normal cats. Am J Vet Res 2000; 61: 1016–1020.
14. Ivancić M, Mai W. Qualitative and quantitative comparison of renal vs hepatic ultrasonographic intensity in healthy dogs. Vet Radiol Ultrasound 2008; 49: 368–373.
15. Zotti A, Banzato T, Gelain ME, et al. Correlation of renal histopathology with renal echogenicity in dogs and cats: an ex-vivo quantitative study. BMC Vet Res 2015; 11: 99.
16. Banzato T, Gelain ME, Aresu L, et al. Quantitative analysis of ultrasonographic images and cytology in relation to histopathology of canine and feline liver: an ex-vivo study. Res Vet Sci 2015; 103: 164–169.
17. Thijssen JM, Herzog K, Weijers G, et al. Ultrasound image analysis offers the opportunity to predict plasma progesterone concentrations in the estrous cycle in cows: a feasibility study. Anim Reprod Sci 2011; 127: 7–15.
18. Banzato T, Fiore E, Morgante M, et al. Texture analysis of B-mode ultrasound images to stage hepatic lipidosis in the dairy cow: a methodological study. Res Vet Sci 2016; 108: 71–75.
19. Starke A, Haudum A, Weijers G, et al. Noninvasive detection of hepatic lipidosis in dairy cows with calibrated ultrasonographic image analysis. J Dairy Sci 2010; 93: 2952–2965.
20. Scully S, Evans ACO, Carter F, et al. Ultrasound monitoring of blood flow and echotexture of the corpus luteum and uterus during early pregnancy of beef heifers. Theriogenology 2015; 83: 449–458.
21. Motta L, Mandara MT, Skerritt GC. Canine and feline intracranial meningiomas: an updated review. Vet J 2012; 192: 153–165.
22. Haralick R, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern 1973; 3: 610–621.
23. Haralick R. Statistical and structural approaches to texture. Proc IEEE 1979; 67: 786–804.
24. Hu Y, Dennis T. Textured image segmentation by context enhanced clustering. IEE Proc Vis Image Signal Process 1994; 141: 413–421.
25. Lerski RA, Straughan K, Shad LR, et al. MR image texture analysis—an approach to tissue characterization. Magn Reson Imaging 1993; 11: 873–887.
26.MaZda user's manual. Lodz, Poland: Technical University of Lodz, Institute of Electronics, 2005. Available at: www.eletel.p.lodz.pl/programy/mazda/download/mazda_manual.pdf. Accessed Mar 20, 2017.
27. Harrison LCV, Luukkaala T, Pertovaara H, et al. Non-Hodgkin lymphoma response evaluation with MR texture classification. J Exp Clin Cancer Res 2009; 28: 87.
28. Szczypicski PM, Strzelecki MM, Materka A, et al. MaZda: a software package for image texture analysis. Comput Methods Programs Biomed 2009; 94: 66–76.
29. Strzelecki MM, Szczypinski PM, Materka A, et al. A software tool for automatic classification and segmentation of 2D/3D medical images. Nucl Instrum Methods Phys Res A 2013; 702: 137–140.
30. Herlidou-Měme S, Constans JM, Carsin B, et al. MR texture analysis on texture test objects, normal brain and intracranial tumors. Magn Reson Imaging 2003; 21: 989–993.
31. Davnall F, Yip CSP, Ljungqvist G, et al. Assessment of tumor heterogeneity: an emerging imaging tool for clinical practice? Insights Imaging 2012; 3: 573–589.
32. Eliat PA, Olivié D, Saïkali S, et al. Can dynamic contrast-enhanced magnetic resonance imaging combined with texture analysis differentiate malignant glioneuronal tumors from other glioblastoma? Neurol Res Int 2012; 2012: 195176.
33. Singh JB, Oevermann A, Lanh J, et al. Contrast media enhancement of intracranial lesions in magnetic resonance imaging does not reflect histopathologic findings consistently. Vet Radiol Ultrasound 2011; 52: 619–626.
34. Keyerleber MA, Mcentee MC, Farrelly J, et al. Three-dimensional conformal radiation therapy alone or in combination with surgery for treatment of canine intracranial meningiomas. Vet Comp Oncol 2015; 13: 385–397.
Appendix 1
Texture features calculated by TA softwarea from MRI scans of the brain and spinal cord in dogs with meningioma.
| Texture feature | Parameters | Description |
|---|---|---|
| Histogram | Mean, variance, skewness, kurtosis, percentiles (1%, 10%, 50%, 90%, 99%) | These features describe the mathematical characteristics (mean, variance, skewness, and kurtosis) of the histogram of pixel intensity values. |
| Gradient-based histogram | GrMean, GrVariance, GrSkewness, GrKurtosis, GrNonZeros | These features describe the gradient of the pixel intensities calculated by use of a 3 × 3-pixel interval. |
| Run-length matrix | Horzl_RLNonUni, Horzl_GLevNonU, Horzl_LngREmph, Horzl_ShrtREmp, Horzl_Fraction, Vertl_RLNonUni, Vertl_GLevNonU, Vertl_LngREmph, Vertl_ShrtREmp, Vertl_Fraction, 45dgr_RLNonUni, 45dgr_GLevNonU, 45dgr_LngREmph, 45dgr_ShrtREmp, 45dgr_Fraction, 135dr_RLNonUni, 135dr_GLevNonU, 135dr_LngREmph, 135dr_ShrtREmp, 135dr_Fraction | These features describe the homogeneity of pixel intensities along specific directions of the ROI. |
| Autoregressive model | Teta1, Teta2, Teta3, Teta4, Sigma | These features evaluate the randomness of signal intensity across the ROI. |
| Wavelet transform | WavEnLL_s-(1–5), WavEnLH_s-(1–5), WavEnHL_s-(1–5) | These features quantify the frequency of similar pixel intensity in the ROI. |
| Co-occurrence matrix | S(x,x)AngScMom, S(x,x)Contrast, S(x,x)Correlat, S(x,x)SumOfSqs, S(x,x)InvDfMom, S(x,x)SumAverg, S(x,x)SumVarnc, S(x,x)SumEntrp, S(x,x)Entropy, S(x,x)DifVarnc, S(x,x)DifEntrp | These features describe the homogeneity of the pixel intensity across the ROI at increasing pixel distances (x ranges from −5 to 5). |
Appendix 2
Top 30 discriminating texture features extracted by TA softwarea from MRI scans of the brain and spinal cord of dogs with meningioma.
| T1-weighted images | Postcontrast T1-weighted images | T2-weighted images |
|---|---|---|
| WavEnHL_s-2 | S(2,0)AngScMom | Perc.99% |
| Teta3 | S(1,–1)Entropy | Perc.10% |
| Teta2 | S(1,–1)AngScMom | Vertl_Fraction |
| S(5,–5)Entropy | S(0,1)Entropy | S(3,0)DifVarnc |
| S(5,5)SumVarnc | S(0,2)Entropy | S(2,2)SumAverg |
| S(0,5)Entropy | S(4,0)AngScMom | S(2,–2)Entropy |
| S(0,5)SumVarnc | S(0,5)SumVarnc | S(0,2)DifEntrp |
| S(5,0)Correlat | S(2,–2)AngScMom | S(3,0)Correlat |
| S(4,–4)SumAverg | S(2,2)AngScMom | Vertl_RLNonUni |
| S(4,4)SumAverg | S(1,1)Entropy | Perc.50% |
| S(0,4)Entropy | S(4,0)Contrast | Teta1 |
| S(0,4)SumEntrp | Perc.10% | S(5,–5)SumAverg |
| S(4,0)Entropy | S(1,0)SumVarnc | S(0,5)DifVarnc |
| S(3,3)Entropy | Teta4 | WavEnHL_s-3 |
| S(3,3)SumVarnc | S(4,4)Entropy | S(0,1)AngScMom |
| S(0,3)Entropy | S(5,–5)SumAverg | Teta3 |
| S(3,0)Entropy | Teta3 | Mean |
| S(2,–2)Entropy | S(4,4)DifVarnc | S(5,0)DifEntrp |
| S(2,–2)SumAverg | S(2,–2)Entropy | S(5,5)SumAverg |
| S(2,2)Entropy | S(4,0)Entropy | Perc.90% |
| S(0,2)Entropy | S(0,5)Entropy | Perc.01% |
| S(1,–1)SumOfSqs | S(2,0)Entropy | WavEnHH_s-3 |
| S(1,1)Entropy | S(3,–3)Entropy | S(5,0)DifVarnc |
| S(1,1)SumVarnc | S(3,3)Entropy | WavEnLL_s-1 |
| S(0,1)Entropy | S(3,0)Entropy | S(0,1)DifVarnc |
| S(1,0)Entropy | S(2,2)Entropy | S(4,0)DifEntrp |
| Perc.99% | S(0,4)Entropy | Variance |
| Perc.90% | S(0,3)Entropy | S(0,1)Contrast |
| Perc.10% | S(0,5)SumEntrp | S(1,1)Correlat |
| Perc.01% | S(5,0)Entropy | S(4,0)DifVarnc |
