Non-Hodgkin lymphoma response evaluation with MRI texture classification

MRI images of nineteen NHL patients participating in the TRE project were selected for the first part of this study. One of these patients was excluded due to the smaller amount of image data from the second part analyses. There were 14 male and 5 female patients aged 34–75. These patients had untreated or relapsed histologically diagnosed high/intermediate (N = 8, 42%) or low-grade (N = 11, 58%) NHL with an evaluable lymphoma lesion either in the abdominal area (N = 16) or in the clavicular and axillary lymph node area (N = 3). The treatment given was chemotherapy alone or combined with humanized antibody, rituximab (Mabthera^®). Therapy regimens were CHOP (N = 5), R-CHOP (rituximab and CHOP) (N = 8), and CVP (cyclophosphamide, vincristine and prednisone) (N = 1), CHOP-like CNOP (cyclophosphamide, mitoxantrone, vincristine and prednisone) (N = 1), ChlP (chlorambucil and prednisone) (N = 1), starting with CHOP and changing to R-CHOP (N = 2), starting with R-CHOP and changing to R-CVP (N = 1). Chemotherapy regimens were selected according to patients' clinical status. Chemotherapy courses were repeated every three weeks, and 4 to 9 courses were given according to clinical response. Two patients received 4 cycles, four patients 6 cycles, one patient 7 cycles, and 11 received 8 cycles, and one 9 cycles.

MR imaging in clinical practice as well as in this study was carried out at staging phase before any treatment (examination 1, E1), after the first chemotherapy cycle (examination 2, E2), and after the fourth chemotherapy cycle (examination 3, E3). In addition patients were followed up by using MRI six months and 6–61 months after the completion of therapy. The time frame of the study is presented in Figure 1.

Imaging was performed on a 1.5 T MRI device (GE Signa, Wisconsin, USA).

One contrast enhanced sequence acquired from the first and second imaging timepoint were included for volume analysis of lymphoma masses. The sequence used was axial T2-weighted fast spin echo (FSE) fat saturation (FAT SAT) sequence (TR 620 ms, TE 10 ms), with intravenous contrast agent gadolinium chelate (gadobenate dimeglumine, 0.2 mg/ml, 10 ml), slice thickness ranged from 5 mm to 12 mm.

One or two T1- and T2-weighted axial image serquences from the first three imaging timepoints of every patient were taken for texture analysis. The T1-weighted series comprised T1-weighted spin echo (SE) and T1-weighted SE FAT SAT sequences (TR 320–700 ms, TE 10 ms), the T2-weighted sequences were FSE FAT SAT (TR 3 320–10 909 ms, TE 96 ms). Repetition time TR varied between and within patients. Slice thickness varied between patients according to clinical status from 5 mm to 12 mm; most patients had two different slice thickness series, the general combination was 5 mm and 8 mm series. Pixel size varied from 1.33 mm*1.33 mm to 1.80 mm*1.80 mm, and a 256*256 matrix was used.

Texture parameter calculation was the first stage of the texture analyses. Stand-alone DICOM viewer application was used to select three to five slices from every image series for analysis. Region of interest (ROI) setting and texture analysis were carried out with MaZda software (MaZda 3.20, The Technical University of Lodz, Institute of Electronics) [33, 34]. The lymphoma masses were manually selected and set as ROIs (Figure 2). Texture features calculated were based on histogram, gradient, run-length matrix, co-occurrence matrix, autoregressive model and wavelet-derived parameters [34]. Image grey level intensity normalization computation separately for each ROI was performed with method limiting image intensities in the range [μ-3σ, μ+3σ], where μ is the mean grey level value and σ the standard deviation. This method has been shown to enhance differences between two classes when comparing image intensity normalization methods in texture classification [35].

Fisher coefficient (Fisher) and classification error probability (POE) combined with average correlation coefficients (ACC) provided by MaZda were used to identify the most significant texture features to discriminate and classify the three evaluation stages of lymphoma tissue. Ten texture features were chosen by both methods (Fisher, POE+ACC). This feature selection was performed separately for the T1- and T2-weighted image sets. In these subgroups feature selection was run for the following imaging stages: combination of all imaging timepoints (E1, E2, and E3), and all combinations of the two aforementioned. Slice thickness was not taken into account.

The volumetry of the solid lymphoma masses was evaluated between diagnostic stage (E1) and after the first treatment (E2). The masses were selected for evaluation before chemotherapy. The same masses were followed after the first treatment. Volumetric analysis based on MRI images was performed with semiautomatic segmentation software Anatomatic™ [36] with region growing method. [37].

The patients' subjective views on their clinical symptoms was observed between two stages: at the diagnosis and after the first treatment. The subjective views were set in two groups: symptoms unchanged or relieved.

Grade of malignity was classed into two groups: 1) low; 2) high/intermediate.

B11 application (version 3.4) of MaZda software package was used for texture data analysis and classification. Analyses were run between all combinations of imaging stages separately for T1- and T2-weighted images. Analyses were performed for combination of parameters selected automatically with Fisher and POE+ACC methods for 1) the specific imaging timepoint pair in question and 2) for all imaging stages in particular image type (T1-, T2-weighted). Feature standardization was used in B11, the mean value being subtracted from each feature and the result divided by the standard deviation. Raw data analysis (RDA), principal component analysis (PCA), and linear (LDA) and nonlinear discriminant analysis (NDA) were run for each subset of images and chosen texture feature groups. B11 default neural network parameters were used. Nearest-neighbor (1-NN) classification was performed for the raw data, the most expressive features resulting from PCA and the most discriminating features resulting from LDA. Nonlinear discriminant analysis carried out the classification of the features by artificial neural network (ANN). These classification procedures were run by B11 automatically.

Statistical analyses were run for the texture features MaZda's automatic methods (Fisher and POE+ACC) had shown to give best discrimination between imaging timepoints. The T1- and T2-weighted image texture parameters were tested separately. Texture parameters for 18 patients were included in the test, one patient participating in MaZda texture parameter calculation was excluded because of smaller amount of image data than other patients leading to reduced textural data.

In analyzing and seeking the best parameters for classification, it is vital to ensure low overall variation in the treatment process and to ascertain how this variation can be focused onto different components in the whole process. In the present study the repeatability and reproducibility (R&R) method was applied. The design of the study was experimental, the aim being to estimate different sources of variation in the lymphoma texture at the three different timepoints (examinations 1, 2, and 3) and repeating the same measurements three times. Because the distributions were skewed, the range method was used.

According to the standard Gage R&R terminology timepoints stand for operators, patients for parts and repeated measurements for trials. In statistical terms the following variance components were estimated: repeatability (difference across measurements), reproducibility (difference across timepoints) and variability (difference across patients). Repeatability describes intrapatient variation, i.e., how a given measurer repeats the same planning process. Reproducibility describes interpatient variation, i.e., how different measurements at the timepoints follow the same planning process and variability describes interpatient variation, i.e. how well the same physician can repeat the planning process for different kinds of patients. The total error – also known as the combined R&R effect – includes repeatability and reproducibility, and only patient-to-patient variation is excluded. In industrial applications the combined R&R should not exceed 10% of the total variation, but in certain situations a total error up to 30% may be acceptable. The present statistical analyses were performed by Statistica/W (Version 5.1, 98 edition, Statsoft. Inc, Tulsa, OK, USA).

Textural data from T1- and T2-weighted fat saturation image series were analysed separately and both groups divided into two subgroups according to slice thickness: 5–7 mm and 8–12 mm. Differences between imaging timepoints were analysed by Wilcoxon Signed Ranks.

Mann-Whitney test was used to test rank parameters grouped by grade of malignity and subjective change of symptoms. These analyses were performed by SPSS for Windows, version 14.0.2.

The median volume of the lymphoma masses before treatment (E1) was 429 cm³, ranging from 72 cm³ to 2144 cm³. The median volume of the masses calculated from the second imaging timepoint (E2) was 190 cm3, ranging from 30 cm³ to1622 cm³. After the first treatment cycle, the lymphoma mass volume had decreased in all patients. The median decline in volume was 32%, ranging from 3% to 76%. The results of this volumetric analysis have been published earlier in more detail [37]. The volumetry results of the first and second imaging are given in cm³, and the volume change is calculated in percentages in Table 1.

According to the patient's subjective estimates clinical symptoms between first and second imaging timepoint were unchanged in eight patients and relieved in 11 patients. Grades of malignancy and subjective view on symptoms are presented in Table 1 with volumetry results.

We included in the analyses 108 T1-weighted and 113 T2-weighted images from E1; 103 T1-weighted and 105 T2-weighted images from E2; and 97 T1-weighted images and 99 T2-weighted images from E3.

Texture features were selected with Fisher and POE+ACC methods in MaZda from 300 original parameters calculated for each of the four subgroups in both image data classes T1- and T2-weighted.

We found that the most significant features varied clearly between imaging stages. The whole of 74 TA features ranked first to tenth significant feature in tested subgroups. There were three histogram parameters, 55 co-occurrence parameters, nine run-length parameters, four absolute gradient parameters and three autoregressive model parameters. No wavelet parameters were placed in the top group.

Data analyses RDA, PCA, LDA and NDA show texture changes between imaging points. The analyses did not perform well the task of discriminating all three imaging timepoints (E1, E2, E3) at same time. Slightly better classification was achieved between the first and second examinations, and between the second and third examinations. The method was successful in classifying the textural data achieved from the pre-treatment and third imaging timepoints, the best discrimination was obtained within T2-weighted leading to NDA classification error of 4%, and within T1-weighted NDA 5% error. Classification of different examination stages lead to same level results in T1- and T2-weighted images. The overall classification results are presented in Table 2 and Table 3.

The values of 73 features obtained with MaZda feature selection methods were tested with Wilcoxon paired test for groups obtained from imaging timepoints a) E1 and E2, b) E2 and E3, c) E1 and E3. T1- and T2-weighted fat saturation image series data were set as their own groups and further into two subgroups according to slice thickness: 5–7 mm and 8–12 mm.

R&R test parameter repeatability was used to describe the variation in texture features between image slices within imaging sequence, and parameter reproducibility to describe the variation between examination stages. This test was performed separately for T1- and T2-weighted images in all three combinations of two imaging points. Differences in slice thickness were not taken into account. Reproducibility values were expected to be quite large because the aim was that the treatment given between imaging stages would take effect and be shown in image texture. In contrast, repeatability values (i.e. differences between images taken at the same timepoint) were expected to be zero. There is no exact expected ratio for reproducibility and patient-to-patient variation in such studies and thus no exact value for percentage of reproducibility, so that the difference between different imaging stages was significant.

The texture parameters giving the best discrimination within T1-weighted image groups in two imaging stage comparison are given in Table 4, Table 5 and Table 6; and respectively for T2-weighted image groups in Table 7, Table 8 and Table 9. Reproducibility percentage and Repeatability percentage of the total are given for all parameters. Wilcoxon paired test p-values are given for all parameters for separate groups regarding slice thickness (groups 5–7 mm and 8–12 mm).

R&R inverted ratio and the small difference between values are associated with poor results in Wilcoxon test with certain exceptions. Comparisons between first and third imaging points achieved significant Wilcoxon test p-values most consistently: within T2-weighted images in both slice thickness groups, and within T1-weighted images in the group of thinner slices. Features ranked in T1-weighted image data were tested in T2-weighted image data and vice versa. These tests with ranked features transposed with T1- and T2-weighted image groups lead to statistically relevant p-values in thinner T1-weighted images and all images in T2-weighted group. In the analyses of first and second imaging timepoints thin slices in general achieved poorer separation than thick slices. Between the second and third imaging sessions Wilcoxon test gave an unsatisfactory result in T1-weighted group. This trend can be seen in the B11 classification results in the framework of T1-weighted images, while the T2-weighted image analyses in B11 show better classification between second and third than first and second imaging points. The best overall discrimination between imaging timepoints in T1-weighted images was given by the run-length matrix parameters describing grey level non-uniformity, run-length non-uniformity, short-run emphasis and fraction of image in runs in one or more directions calculated (horizontal, vertical, 45 degrees and 135 degrees). In the framework of T2-weighted image analyses best the performers were absolute gradient mean and grey level non-uniformity There were some scattering in well acquitted parameters between sub analyses.

Mann-Whitney test was performed for all texture features ranked 1–5 in any classification sub-analysis separately in T1- and T2-weighted images and further subgroups according to slice thickness to analyze differences between stage of malignity (low vs. high/intermediate) and between subjective change of symptoms (unchanged vs. relieved). These analyses did not yield any relevant and consequential additional information on the relation of texture features to grouping parameters.