SUPPLEMENTARY APPENDIX A: Definition of texture features.

Transcription

1 SUPPLEMETARY APPEDIX A: Definition of texture features. Input volume: Volume of interest V (x, y, z) with isotropic voxel size. The necessity for isotropically resampling V to a given voxel size prior to texture analysis results from the fact that Global features are computed from histograms counting the number of gray-levels in 3D space, and that all higher-order texture measurements explicitly or implicitly involve a distance parameter in the matrix computation (GLCM, GLRLM, GLSZM and GTDM). For this parameter to be meaningful in 3D space and in order for the orientation dependence of the tumour to be minimized, isotropic resolution is required. Global texture features (first-order gray-level statistics). Let P define the first-order histogram of a volume V (x, y, z) with isotropic voxel size. P (i) represents the number of voxels with gray-level i, and represents the number of gray-level bins set for P. The i th entry of the normalized histogram is then defined as: p(i) = P (i) P (i). The Global texture features are then defined as: Variance: Skewness: Kurtosis σ 2 = (i µ) 2 p(i) g s = σ 3 (i µ) 3 p(i) g k = σ 4 [ (i µ) 4 p(i) ] 3 Gray-Level Co-occurence Matrix (GLCM) texture features. Let P define the GLCM of a quantized volume V (x, y, z) with isotropic voxel size. P (i, j) represents the number of times voxels of gray-level i were neighbours with voxels of gray-level j in V, and represents the pre-defined number of quantized gray-levels set in V. Only one GLCM of size is computed

2 per volume V by simultaneously adding up the frequency of co-occurences of all voxels with their 26-connected neighbours in 3D space, with all voxels (including the peripheral region) considered once as a center voxel (according to [], thus always using d = ). To account for discretization length differences, neighbours at a distance of 3 voxels around a center voxel increment the GLCM by a value of 3, neighbours at a distance of 2 voxels around a center voxel increment the GLCM by a value of 2, and neighbours at a distance of voxel around a center voxel increment the GLCM by a value of. The entry (i, j) of the of the normalized GLCM is then defined as: = The following quantities are also defined: P (i, j) P (i, j). µ i = i, µ j = j, g g σ i = (i µ i ) 2, σ j = (j µ j ) 2. The GLCM texture features are then defined as: Energy []: Contrast []: energy = [] 2 contrast = (i j) 2 Correlation (adapted from []): correlation = Homogeneity (adapted from []): homogeneity = (i µ i ) (j µ j ) σ i σ j + i j 2

3 Variance (adapted from []): [ variance = (i µi ) 2 + (j µ j ) 2 ] Sum Average (adapted from []): g sum average = [i + j ] Entropy []: ( ) entropy = log 2 Gray-Level Run-Length Matrix (GLRLM) texture features. Let P define the GLRLM of a quantized volume V (x, y, z) with isotropic voxel size. P (i, j) represents the number of runs of gray-level i and of length j in V, represents the pre-defined number of quantized gray-levels set in V, and L r represents the length of the longest run (of any gray-level) in V. Only one GLRLM of size L r is computed per volume V by simultaneously adding up all possible longest run-lengths in the 3 directions of 3D space (one voxel can be part of multiple runs in different directions, but can be part of only one run in a given direction). A MATLAB toolbox created by Xunkai Wei [7] computes GLRLMs from 2D images, and it can be used to facilitate the computation of GLRLMs from 3D volumes. To account for discretization length differences, runs constructed from voxels separated by a distance of 3 increment the GLRLM by a value of 3, runs constructed from voxels separated by a distance of 2 increment the GLRLM by a value of 2, and runs constructed from voxels separated by a distance of increment the GLRLM by a value of. The entry (i, j) of the of the normalized GLRLM is then defined as: = The following quantities are also defined: P (i, j) Lr P (i, j). L r µ i = i, µ j = j. The GLRLM texture features are then defined as: 3

4 Short Run Emphasis (SRE) [2]: Long Run Emphasis (LRE) [2]: SRE = j 2 LRE = j 2 Gray-Level onuniformity (GL) (adapted from [2]): g L r GL = Run-Length onuniformity (RL) (adapted from [2]): L r g RL = Run Percentage (RP) (adapted from [2]): RP = Lr Lr j Low Gray-Level Run Emphasis (LGRE) [4]: LGRE = High Gray-Level Run Emphasis (HGRE) [4]: i 2 HGRE = i 2 Short Run Low Gray-Level Emphasis (SRLGE) [5]: SRLGE = i 2 j

5 Short Run High Gray-Level Emphasis (SRHGE) [5]: SRHGE = i 2 j 2 Long Run Low Gray-Level Emphasis (LRLGE) [5]: LRLGE = j 2 i 2 Long Run High Gray-Level Emphasis (LRHGE) [5]: LRHGE = i 2 j 2 Gray-Level Variance (GLV) (adapted from [6]): GLV = (i µ i ) 2 L r Run-Length Variance (RLV) (adapted from [6]): RLV = (j µ j ) 2 L r Gray-Level Size Zone Matrix (GLSZM) texture features. Let P define the GLSZM of a quantized volume V (x, y, z) with isotropic voxel size. P (i, j) represents the number of 3D zones of gray-levels i and of size j in V, represents the pre-defined number of quantized gray-levels set in V, and L z represents the size of the largest zone (of any gray-level) in V. One GLSZM of size L z is computed per volume V by adding up all possible largest zone-sizes, with zones constructed from 26-connected neighbours of the same gray-level in 3D space (one voxel can be part of only one zone). The entry (i, j) of the normalized GLSZM is then defined as: = The following quantities are also defined: P (i, j) Lz P (i, j). L z µ i = i, µ j = j. 5

6 The GLSZM texture features are then defined as: Small Zone Emphasis (SZE) [2, 6]: SZE = Large Zone Emphasis (LZE) [2, 6]: j 2 LZE = j 2 Gray-Level onuniformity (GL) (adapted from [2, 6]): g L z GL = Zone-Size onuniformity (ZS) (adapted from [2, 6]): Zone Percentage (RP) [2, 6]: L z g ZS = ZP = Lz Lz j Low Gray-Level Zone Emphasis (LGZE) [4, 6]: LGZE = i 2 High Gray-Level Zone Emphasis (HGZE) [4, 6]: HGZE = i

7 Small Zone Low Gray-Level Emphasis (SZLGE) [5, 6]: SZLGE = i 2 j 2 Small Zone High Gray-Level Emphasis (SZHGE) [5, 6]: SZHGE = i 2 j 2 Large Zone Low Gray-Level Emphasis (LZLGE) [5, 6]: LZLGE = j 2 i 2 Large Zone High Gray-Level Emphasis (LZHGE) [5, 6]: LZHGE = i 2 j 2 Gray-Level Variance (GLV) (adapted from [6]): GLV = (i µ i ) 2 L z Zone-Size Variance (ZSV) (adapted from [6]): ZSV = (j µ j ) 2 L z eighbourhood Gray-Tone Difference Matrix (GTDM) texture features. Let P (i) define the GTDM of a quantized volume V (x, y, z) with isotropic voxel size. P (i) represents the summation of the gray-level differences between all voxels with gray-level i and the average gray-level of their 26-connected neighbours in 3D space. represents the pre-defined number of quantized gray-levels set in V, and ( ) eff is the effective number of gray-levels in V, with ( ) eff < (let the vector of gray-levels values in V be denoted as g = g(), g(2),..., g( ); some gray-levels excluding g() and g( ) may not appear in V due to different quantization schemes). One GTDM of size 7

8 is computed per volume V. To account for discretization length differences, all averages around a center voxel located at position (j, k, l) in V are performed such that the neighbours at a distance of 3 voxels are given a weight of / 3, the neighbours at a distance of 2 voxels are given a weight of / 2, and the neighbours at a distance of voxel are given a weight of. The i th entry of the GTDM is then defined as: { all voxels {i} P (i) = i A i if i > 0, 0 if i = 0. where { i } is the set of all voxels with gray-level i in V (including the peripheral region), i is the number of voxels with gray-level i in V, and A i is the average gray-level of the 26-connected neighbours around a center voxel with gray-level i and located at position (j, k, l) in V such that: A i = A(j, k, l) = m= n= m= o= n= m= m= o= w m,n,o V (j + m, k + n, l + o) n= o= n= o= w, m,n,o if j m + k n + l o =, where w m,n,o = 2 if j m + k n + l o = 2, 3 if j m + k n + l o = 3, 0 if V (j + m, k + n, l + o) is undefined. The following quantity is also defined: n i = i. where is the total number of voxels in V. The GTDM texture features are then defined as: Coarseness [3]: g coarseness = ɛ + n i P (i) where ɛ is a small number to prevent coarseness becoming infinite. Contrast [3]: contrast = [ ( ) eff (g ) eff ] g n i n j (i j) 2 g P (i) Busyness [3]: busyness = n i P (i) (i n i j n j ), n i 0, n j 0 8

9 Complexity [3]: g i j [ n i P (i) + n j P (j) ] complexity =, n i 0, n j 0 (n i + n j ) Strength [3]: strength = (n i + n j ) (i j) 2 [ ɛ + ], n i 0, n j 0 g P (i) where ɛ is a small number to prevent strength becoming infinite. Online ressources MATLAB R codes are available at: References [] Haralick, R.M., Shanmugam, K. and Dinstein, I. (973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, smc 3(6), [2] Galloway, M.M. (975). Texture analysis using gray level run lengths. Computer Graphics and Image Processing, 4(2), [3] Amadasun, M. and King, R. (989). Textural features corresponding to textural properties. IEEE Transactions on Systems, Man, and Cybernetics, 9(5), [4] Chu, A. Sehgal, C.M. and Greenleaf, J.F. (990). Use of gray value distribution of run lengths for texture analysis. Pattern Recognition Letters, (6), [5] Dasarathy, B.V. and Holder, E.B. (99). Image characterization based on joint gray level-run length distributions. Pattern Recognition Letters, 2(8), [6] Thibault, G. et al. (2009). Texture indexes and gray level size zone matrix. Application to cell nuclei classification. In Pattern Recognition and Information Processing. Minsk, Belarus, [7] Wei, X. (2007). Gray Level Run Length Matrix Toolbox v.0, computer software. Beijing Aeronautical Technology Research Center. Available from: [ January 203] 9

10 SUPPLEMETARY APPEDIX B: Texture features with highest prognostic value. Table B.. Texture features with the highest Spearman s rank correlation (r s) with lung metastases in STS cancer and their associated extraction parameters, for the five types of scans investigated in this study. The values in italic font indicate features for which p < 0.05/ SCA TEXTURE rs MRI Inv. MRI weight. WPBF Scale Quant. algo. g FDG-PET RLV 0.57 R = 2/3 mm Lloyd-Max 6 T SZE 0.5 R = 2/3 5 mm Lloyd-Max 64 T2FS LRLGE 0.40 R = 2/3 2 mm Equal 64 FDG-PET/T SZE 3 Inv. 3/4 R = 2/3 5 mm Lloyd-Max 64 FDG-PET/T2FS SZE 0 o Inv. /2 R = 3/2 3 mm Lloyd-Max 64 MRI Inv.: MRI Inversion, MRI weight: weight in FDG-PET/MRI fusion process, WBPF : wavelet band-pass filtering, Scale: isotropic voxel size, Quant. algo.: quantization algorithm, g : umber of gray-levels, R: ratio of weight given to band-pass sub-bands to weight given to low- and high-frequency sub-bands. 0

11 SUPPLEMETARY APPEDIX C: Degrees of freedom on texture extraction parameters. Baseline Full Optimal FDG PET SCAS SEPARATE SCAS FUSED SCAS AUC SESITIVITY SPECIFICITY Model Order Figure C.. Estimation of prediction performance as a function of model order for multivariable models constructed from the FDG-PET, SEPARATE, and FUSED scans for three types of experiments: i) experiments in which no texture extraction parameter was allowed to vary (baseline parameters); 2) experiments in which all texture extraction parameters were allowed to vary (full degree of freedom); 3) experiments using the particular set of texture extraction parameters allowed to vary yielding the highest 32+ bootstrap AUC, separately for each model order of each initial feature set (optimal degree of freedom). Error bars represent the standard error of the mean on a 95 % confidence interval. There exists a certain degree of model complexity for which prediction models have optimal parsimonious properties. An additional factor to model order that increases model complexity comes from the whole range of texture extraction parameters used in this work, as too specific extraction parameters may reduce the generalizability of prediction models. One way of reducing model complexity is to fix specific extraction parameters by preventing them to fully vary, or in other words, to restrict the degree of freedom on texture extraction parameters. In principle, higher-order models should be coupled to more restricted extraction parameter degrees of freedom in order to reach optimal parsimonious properties, and vice-versa. The results presented in Figure C. confirm that there exists an optimal combination of model order and texture extraction parameter degree of freedom in terms of predictive properties. It can be seen that the higher the model order, the more distance there is between the curves obtained using full and optimal degree of freedom on texture extraction parameters, especially in the case of the FDG-PET feature set. On the other

12 hand, for lower model orders, the curves obtained using full and optimal degree of freedom on texture extraction parameters are generally identical. Table C. shows which extraction parameters were allowed to vary in the experiments yielding the optimal prediction performance estimation in Figure C., for model orders of to 0 and for the three different types of initial feature sets. Table C.. Texture extraction parameters allowed to vary in the experiments yielding optimal prediction performance estimation. It can now be directly seen from Table C. that an optimal degree of freedom on texture extraction parameters generally allows less parameters to vary for higher model orders, except in the case of the feature set using separate scans. However, in light of the results we obtained, it is not possible to identify a trend about which texture extraction parameters should be used as a function of model order to obtain optimal prediction performance estimation. 2

13 SUPPLEMETARY APPEDIX D: Uncertainty analysis. D.. Contouring variations. In this section, an estimation of the uncertainty of the multivariable model proposed in this work g(x i ) due to contouring variations is evaluated. Thirty-two patients of the cohort had visible edema that could be clearly identified in the vicinity of their tumours. For these patients, two types of contours were drawn: one incorporating the visible edema (denoted as Edema), and one excluding it (denoted as Mass). The proposed model and its associated logistic regression coefficients were obtained using textures computed with the Mass contours, and we now evaluate how its response changes using textures computed with the Edema contours (but with the same logistic regression coefficients). Table D. first presents the volumes of the tumours and the textures features differences of the four features of the proposed model extracted with the two contours, for all patients of the cohort. In table D., the average x j of the absolute difference between the values x ij of the four features j =, 2, 3, 4 of the proposed model extracted from the Mass and Edema contours over all patients i =, 2,..., was computed such that: x ij = x ij [Edema] x ij [Mass], for j =, 2, 3, 4 x j = x ij, for j =, 2, 3, 4 For the 9 patients without the Edema contour, x j was considered to be 0 and was still incorporated in the calculation of x j. Then, the global contribution of contouring variations on the uncertainty of g(x i ) was obtained by adding in quadrature the contribution of each average texture variation x j such that(recalling that g(x i ) = 256 x i x i x i x i ): [ p ( g(xi ) 2 ɛ contour = ) ( x ] ) 2 j x ij ɛ contour = ( 256) 2 (0.0065) 2 + (5360) 2 ( ) 2 + (.75) 2 (.46) 2 + (3.6) 2 (0.39) 2 ɛ contour 4.89 Overall in our patient cohort, an absolute value of ±4.89 was calculated for ɛ contour. This uncertainty is constant across all values of g(x i ). 3

14 Table D.. Volumes, texture features and texture features differences of the four features forming the multivariable model proposed in this work in the case where the tumour region was outlined with the contour excluding surrounding edema (Mass), and the case where the tumour was outlined with the contour including the surrounding edema (Edema). Patient Volume Volume Volume Feature Feature Feature Feature Feature Feature Feature Feature number (cm 3 ) (cm 3 ) (cm 3 ) xi x i x i x i2 x i2 x i2 x i3 x i3 x i3 x i4 x i4 x i4 (i) Mass Edema Mass Edema Mass Edema Mass Edema Mass Edema x i : FDG-PET/T2FS - - GLSZM/SZE x i2 : FDG-PET/T - - GLSZM/ZSV x i3 : FDG-PET/T - - GLSZM/HGZE x i4 : FDG-PET/T2FS - - GLRLM/HGRE 4

15 D.2. Bootstrap confidence intervals. Table D.2. Response of the multivariable model proposed in this work on the entire patient cohort, and its associated 95 % confidence interval (CI) bounds. Patient number Model response CI CI Outcome (i) g(x i ) Lower bound Upper bound CI

16 SUPPLEMETARY APPEDIX E: Permutation tests. 30 AUC SESITIVITY SPECIFICITY Performance metrics estimates Figure E.. Distributions of performance metrics estimates of multivariable models of order 4 found in 000 permutation tests using shuffled outcome vectors. All histograms were constructed with 00 bins. 6