Patient Population: In this study,
MR images of 25 healthy individuals ( with average age of 51 years and standard deviation of 17 years) and 25 patient with myeloma (with average age of 56 years and standard deviation of 11 years) had been used.
Out of total 50 patient 18 were male and 32 were female.
Data Acquisition: All MR images were acquired using a 1.5 Tesla system (Achieva,
Philips Medical System). T1,
T2 and diffusion sequences (b values 600 and 1200 s/mm2) were used in this study.
Data preparation and preprocessing:
Flow chart of experiment.
Fig. 1: Flow chart of the experiment.
All MR images were obtained from local hospital with appropriate ethical clearance and anonymization. All images are converted to Nifti format for further analysis.
Ground truth were generated separately from T2 images and DW images with the help of expert rater. T1 images were rigidly registered with respect to T2 images for purpose of using same ground truth of T2.
Min-max normalization was carried out on all images.
Regions of interest were created by multiplying images and their corresponding ground truth.
Only 5 slices,
which have maximum pelvic information,
were used for processing.
Fig. 2: T1-weighted axial image of pelvis of a patient.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 3: T2-weighted axial image of pelvis of a patient.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 4: Mask generated on T2-weighted image of a patient.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 5: DW Image (b value 600) of patient with myloma.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 6: DW Image (b value 1200) of patient with myloma.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 7: Mask generated on DW image of a patient.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 8: T1-weighted axial image of pelvis of a healthy individual.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 9: T2-weighted axial image of pelvis of a healthy individual.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 10: Mask generated on T2-weighted image of a healthy individual.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 11: DW Image (b value 600) of a healthy individual.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 12: DW Image (b value 1200) of a healthy individual.
References: Apollo Cancer Institute and Specialty Hospital.
Fig. 13: Mask generated on DW Image of a healthy individual.
References: Apollo Cancer Institute and Specialty Hospital.
Data Analysis:
Segmented pelvic bone was used to extract different types of features.
These features were classified in to two categories namely first order features (19 features) and texture based features (59 features).
Texture based features were of four types such as Gray Level Co-occurrence Matrix (GLCM) features (27 features) [6],
Gray-Level Run-Length matrix (GLRLM) features(16 features) [7],
Gray Level Size Zone (GLSZM) features(16 features) [8,
9,
10].
15 Feature matrices were created by combining features obtained from different sequence images such as T1 alone,
DWI- 600 alone,
T1 and T2,
All combination,
etc.
To get optimal number of features for feature selection method was applied in two steps.
In first step,
Principal Component Analysis (PCA) was performed on feature matrix and Recursive Feature Elimination with Cross Validation (RFECV) were applied as a final step to select optimal number of features [11].
Several type of classification algorithm such as Support Vector Machine (SVM),
Multi Layer Perceptron (MLP),
Naive bias,
discriminant analysis,
decision tree,
and ensemble classifier were applied on selected feature matrix to classify healthy individual and patient with MM [12].
SVM with Linear kernel (Linear SVM) and with Radial Basis Function kernel (RBF SVM) were used as two variant of SVM.
Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) were used in category of discriminant analysis.
MLP classifier with different activations function (RELU,
Identity and tanh) were used.
To evaluate the performance of classifiers confusion matrix were drawn on test set.
Accuracy,
sensitivity,
specificity,
precision,
F1-measure were calculated for each classifier.
Fig. 14: Confusion Matrix and different metrics.
F1-measure is the harmonic mean of precision and sensitivity.
Accuracy = (True Positve + True Negative)/All