Validating The Discriminating Efficacy Of MR T2 Relaxation Value Of Different Brain Lesions And Comparison With Other Differentiating Factors: Use Of Artificial Neural Network And Principal Component Analysis
T K Biswas, R Bandopadhyay, A Dutta
adc value, ann, metabolites, pca, t2 value
T K Biswas, R Bandopadhyay, A Dutta. Validating The Discriminating Efficacy Of MR T2 Relaxation Value Of Different Brain Lesions And Comparison With Other Differentiating Factors: Use Of Artificial Neural Network And Principal Component Analysis. The Internet Journal of Radiology. 2017 Volume 20 Number 1.
PURPOSE: 1.To judge the effectiveness of MR T2 value determined by T2 relaxation mapping and false color encrypted portray or mapping of the T2 W input image.
2. To estimate the tissue water content of the lesion and perilesional region.
3. To live predict the diseases by Artificial Neural Network (ANN) and clustering by Principal Component Analysis(PCA).
MATERIALS AND METHODS: A T2 relaxation map was prepared in the MR scanner of a normal patient and patients of different grades of glioma, cysts, multiple sclerosis and metastasis and an individual pixel T2 value was determined. A false color shade was prepared depending upon the T2 value to make a color encrypted plan or portray of T2 values overlay on T2 W image. A relationship of T2 value and accumulated tissue water content was established. Quantity of the brain metabolites and apparent diffusion coefficient (ADC) were determined and tabulated. ANN was tried to live predict from these data as input in the excel spread sheet. Clustering of the diseases could be done as well.
CONCLUSION: Discrimination of the diseases by color coded T2 map and live prediction can be done successfully.
MOTIVATION AND BACKGROUND
Accurate diagnosis is needed for the life saving treatment of brain tumors and various other brain diseases. Diagnostic accuracy of CT scan, MRI, PET-CT scan is around 50 to 70% if we compare it with the histopathological diagnosis considered as gold standard (1). MR spectroscopy is another method by which almost 80 % of accurate diagnosis can be made. To avoid the hazards of stereoscopic biopsy (2) there is a real need for noninvasive tissue discriminatory method for the management or treatment of the patients. Combination of data (T2 value, ADC value) of MRI and MR spectroscopy (chemical metabolites) synchronizing with Artificial Neural Network may allow the concomitant assessment of morphologic, metabolic, and molecular information on the human brain and pathological lesions (3).
1. To validate the differentiating effectiveness of T2 value of different brain lesions by T2 relaxation mapping (T2 mapping) (3, 4) and creating color coded T2 mapping of the lesion. Benign and malignant tissues, intra and extra lesional fluid accumulation can be discriminated.
2. To quantify the water content within the solid and peripheral part of the lesions.
3. Considering the important physical and chemical properties of brain tissue, an idea of amalgamating physical properties such as T2 value and Apparent Diffusion Coefficient (ADC) values with the chemical properties such as brain metabolites (determined by MR Spectroscopy) was initiated and assigned with the pixel Gray Scale value (GSV) of MR images to get “ ground truth images”. The physical and chemical factors can be used as independent and dependent variables by different statistical methods such as Artificial Neural Net Work
( ANN) and to judge the accuracy of T2 value and to live predict the diseases (4). Principal Component Analysis (PCA) was used to cluster the discriminating factors and diseases.
PHYSICAL AND CHEMICAL STATE OF BRAIN TISSUE
A. Physical State
a) T2 value of Brain tissue
b) Apparent Diffusion Coefficient (ADC) value
B. CHEMICAL STATE
MR spectroscopic (MRS) pattern graph along with quantification of the metabolites like
NAA, creatine, choline, lactate, lipid and MI (myoinositol) peaks signifies the chemical environment of normal brain and neurological diseases.
C. EVALUATION BY STATISTICAL METHODS AND LIVE PREDICTION BY DATA ANALYTICAL TECHNIQUE
a) Principal Component Analysis (PCA): Clustering of different diseases and various parameters or data collected by experimentation and investigation can be completed by PCA.
b) Artificial Neural Network (ANN): ANN can live predict the diseases and tissue character if various data and results are properly trained and tested.
T2 RELAXATION VALUE OR T2 VALUE AND BRAIN
Water content and solid component of brain tissue is unswervingly associated with the T2 value (5). If it is considered that “A” be the percentage of H2O occuring in the brain tissue, then (100-A) is the fraction of solid component in the tissue. The brain tissue resembles hydro-gel containing water, protein and phospho-lipids in different percentages. Normal water content and T2 value of the CSF and brain tissues like Gray and white matters are depicted in the Table1.1(6,7)
A connection between the water content and relaxation value of T2 of brain tissue was established to check out the tissue water content in the solid part of the lesion and as well as in the perilesional and peritumoral edema.
GENERATION OF T2 MAP
The sensitivity of MRI with different percentage of water molecule in the brain tissue is well appreciated. By using theT2 value, water content and relative solid component of a brain tissue in normal and different clinical condition could be deduced (6, 7, 8). A multi echo read out train was considered to determine the T2 decay and to generate a T2 map from the following equation:
S =S0 e-TE/T2
A T2 map was produced in a 3T magnet to evaluate the T2 values in the image and to show a relationship with the T2 shade (in gray scale) created in the scanner (9).
ADC values can be measured by the software of the scanner automatically as a parametric (ADC) map display (10, 11). From the ADC map degree of diffusion of water particles in the interstitial and intracellular portion of brain tissues can be determined. By placing a ROI, ADC value can be measured from the ADC map (in units of mm2/ sec) (12) (Figure3.3). There is controversy of ADC values with their range of diffusion. It is observed that >1.0 to 1.02 (x 10-3 mm2/s ) can be regarded as restriction of diffusion indicating probability of underlying malignant pathology (10,12)( Figure 1.1).
ADC values attempt to help assessing the clinical, physiological as well as pathological state of the tissue.
Assessment of the quantity or extent of the metabolites derived as MRS signal peaks can indicate the tissue state of the signal creating molecules. Inside the fluid compartments of different biological tissues these low molecular weight materials are easily identifiable as MRS signals and can travel freely. Lipid, myo-inositol, lactate NAA, choline, creatine, glutamine glutamate are the main metabolites (13) (Figure1.2 and Table 3.5). The creatine peak and its concentration in the brain tissue usually remain constant both in normal and pathological state. For this reason creatine can be taken as a reference. Any deviation in the concentration can generate a Specific Pattern Graph (14). Generally, concentration of choline is high and NAA is low in malignant brain tumor (Figure1.2B). In demyelinating disease like multiple sclerosis (MS) choline peak of tissue content is high. In tissue damaging or degenerating state lipid and lactate peak is high (15).
STATISTICAL METHOD FOR CLUSTERING AND DATA ANALYTICAL METHOD FOR LIVE PREDICTION OF DISEASES AND CHARECTERIZATION OF TISSUES
Statistical procedures can help clustering the parameters and disease as well when they are overlapped with each other. The procedures considered in this work are:
a) PRINCIPAL COMPONENT ANALYSIS (PCA)
b) Data analytical method like ARTIFICIAL NEURAL NETWORK (ANN)
In this study preliminarily ADC values, T2 value, and chemical components of the disease were observed from the PCA plots. A supervised neural network model was used to determine the capability of these factors to predict a specific disease. The evaluation was carried out using the 10-fold cross-validation method with the back-propagation model.
To analyze various factors of a dataset and to prepare special clustering or grouping of a particular prototypes or pattern, PCA is required. Its main use is to prepare data straightforward to investigate the insight. PCA is required to prepare one dimension instead of multi dimension. (16,17).
ANN learns the relation between input and output from a training dataset. This technique is generally used to study the correlation between two features. Back-propagation multilayer perception (18,19) can be implemented to live predict possible pathological condition from the data retrieved from MRI as “input”.
The usual procedures were adopted to manipulate neural networks to live forecast the disease with unknown tissues.
The important steps: selected are:
1. To create a Data Set Manager putting all the data in an Excel spread sheet.
2. First training of all the data (parameters) including independent and dependent variables.
3. Testing of the data for prediction.
4. Prediction of trained data.
MATERIALS AND METHOD
The following data are required for tissue mapping and color coding.
i) T2 relaxation value of tissue from T2 mapping generated by a MR machine.
ii) Estimation of water content within the neoplasm or mass and in the periphery involving the surrounding white matter.
ii) ADC Values are determined from the program available in the scanner.
In a 3 Tesla GE Signa HDxt (USA) about 98 both male and female patients of different age group( from 11 to 75 years) were scanned (Table 2.1) after getting proper institutional ethical approval.
MR IMAGE (MRI)
The patients were scanned routinely in a 3Tesla GE Sigma HDXT (USA) MRI machine with sequences like T1 Plain axial, coronal and sagittal, FLAIR, T2, and DWI. Post gadolinium MRI was also considered in some case using Fat saturated T1 sequences (TR: 400 - 500, TE: 8, slice thickness: 5-6 mm) using gadolinium at the dose 0.1 mmol/kg body weight. Taking the b value= 0 (TR: 3350 to 3500, TE: 95 - 120) DW (Diffusion weighted) images can be obtained.
T2 VALUE DETERMINATION
1) T2 map generation. Determination of T2 value of CSF and gray and white matter.
2) To estimate water content;
3) Proportion of solid component of the tissue.
Determination of T2 value from T2 decay:
Different six echo times of 30, 60, 90, 120, 150 and 180 ms (20) and TR 1410 ms were applied to generate a T2 relaxation map in a 3T GE magnet. A head coil (quadrature type) was used to obtain a T2 weighted scan. A matrix of 256 X 256 with a FOV of 210-220 mm and thickness of Slice aboout 10 mm have been selected as parameters at the level of ventricles. ROI of 5 to 6 mm2 was selected to determine the T2 intensity (S) of CSF in the body of the lateral ventricles, gray matter in the parietal lobe, and white matter in the centrum semiovale (21). Appropriate signal strength (S) (retrieved from separate echo time) was utilized to fit into a solitary decay of exponential, S=S0e-TE/T2 , so that a T2 map can be created on a pixel by pixel basis.
T2 values of tissue of a certain pixel can be measured from the T2 map easily (Figure 2.1A) from the ROI. A T2 shade (in gray scale) was also prepared from the machine to get the T2 value directly (Figure 2.1 B).
The T2 values of gray as well as white matter and tissue of other pathological lesions were depicted (Table 3.1) and compared. From the T2 map, the pixel T2 value at that location (having X and Y coordinates) of brain tissue or any lesion (22) can be determined as well. A T2 shade in gray scale was generated as well according to the relaxation values (Figure2.3). A palette of false color-coded shade (Figure 2.3C) was also created to generate a false-color encrypted T2 portray or map superimposed on a background of the T2 weighted “input image” (Figure2.2). Dark and bright red shades were selected for the T2 values (ranging from 50 to 150 ms)
indicating the malignant portion of the tumor. Fluid is denoted by bluish or bluish white shade having long T2 value. The nature of the central and peripheral portion of the mass or lesion and to demonstrate the related perilesional edema thus can be demarcated by the bluish color.
The Pixel T2 values of the a) central part of the solid lesion b) in the periphery about 5 mm and c) 10 mm from the central part of the solid lesion were determined and tabulated to establish a relationship between T2 value and tissue water content. Ultimately, to quantify the water content within the lesion and away from the lesion a relationship was established. This will be discussed in detail in the result section.
PREPARATION OF THE PALETTE (COLOR CODED T2 SHADE)
For the display purpose a T2 shade was prepared in the form of a palette (gray or false color coded) from the T2 values derived from the T2 map (Figure 2.3). The shade is made up of 16 slabs with increment of 6 pixel value of gray scale (0 to 255) or 16X 16=256 total pixel gray value (of eight bit). The darkest lower slab reads T2 value of 50 ms and the brightest upper slab reads T2 value about 350 ms. From the T2 shade a false color shade is prepared as well. The bright red slab denoting T2 value is between 50 to 150 ms symbolizing the malignant portion of the lesion or tumor. The upper blue shade has a T2 value of 275 to 350 ms representing the long T2 value produced by fluid or water.
This palette (Figure2.3) can be incorporated within the Application program of the DICOM editor. Different Palettes or Color Schemes could be created from the CTCS chromatizer program (COLOR CONVERTER ©) (Figure 2.4). There is one gray component (values of 0 to 255) and corresponding color component (color space) which comprises of red, green and blue Scale setting (0 to 256 shades).
An approximate gray scale value and corresponding disease is mentioned in the Table 4.2 in the result and discussion section.
QUANTIFICATION OF THE CHEMICAL METABOLITES BY MR SPECTROSCOPY
By utilizing a quadrature head coil on a 3 Tesla magnet (GE Signetom HDTx,USA) data of 1H MRS were retrieved regularly for the suspected pathological and normal cases. A “10 mm x 10 mm x10 mm” voxel size was selected with TR- 9600 and TE between 110 to 144 ms to conduct single or multi voxel spectroscopy. To note the lactate peak a TE of 35 ms was used frequently. To detect abnormal metabolic peak PRESS or single voxel point-resolved MR spectroscopy using short TE (30 ms) or prolonged TE (135 ms) was performed (22, 23) corresponding with the structural changes or morbid anatomy of the mass or neoplasm in the MR image.
Quantities of choline, NAA, MI, choline-creatine, choline-NAA, creatine-NAA ratio were collected (Figure 2.6). Creatine was considered as the reference peak (24, 25, 26,27). Metabolites were tabulated and are mentioned in the result section (Table3.5).
ADC MAP GENERATION
Generation of the ADC map (both gray and color coded) could be done. The ADC value can be determined by placing the region of interest in the map (Figure3.3) and viewing the look up table. ADC values were tabulated.
STATISTICAL AND DATA ANALYTICAL METHODS TO CLASSIFY OR PREDICT DISEASES:
The PCA transforms (10,11) the data into a dimensionally reduced format with a minimum loss of information. The transformation projects the multi-dimensional data into coordinates that maximize the variance and minimize the correlation in the dataset. It forms clusters of similar classes of data. The program was coded in the X L STAT, the statistical software (France).
Neural Tool 7.5 (Palisade Inc.), a Probabilistic Neural Network (PNN) method of non-linear type was applied for live prediction (28). The neural tool helped to live predict the probable underlying pathological lesions from the different input data such as T2 values, ADC values, metabolites and their ratio obtained from MRI and MR spectroscopy.
Independent variables (as inputs):
Inputs of ANN such as T2 values (central and peripheral part of the lesion), ADC values, quantified values related to chemical metabolites (of MRS) choline, creatine, N-acetyl aspertate (NAA), choline-creatine and choline NAA ratio, lipid, lactate, and myoinositol were considered as independent variables.
A network was created first
Then the network was configured
The net initialized the weights and biases,
Network was then trained.
Network was then tested and validated.,
Network was then in a position to live predict values of unknown tissue or diseases as dependent variables (19).
A “data set manager” to be created in Microsoft Excel with the software Neural Tool ( Palisade Inc, artificial neural networks) containing 10 variables, each cell having 32 cells/ variables containing distinguished or known data. During testing 20% error was allowed as “residual”. If the error was less than 20% a good prediction was achieved (Table 2.2).
WATER CONTENT AND T2 VALUE
T2 value of different tissues determined by T2 mapping are depicted in the Table 3.1. The relationship between T2 of brain tissue and water content along with CSF as internal references was also established.
A relationship between T2 value and water content of the tissue and CSF was established from the table 3.1 as:
Water content = 0.075x+68.481
A connection between the T2 value and water content of tissue (fraction of water) was noted. The contribution of free and bound state of tissue water (29,30) toward producing the image
(Figure 2.5) can be considered. Long T2 value is due to the “free phase” of water. About 20% of the total tissue of water is” bound” or motion restricted (7) and is invisible in the MRI (30). In both malignant and benign tumor, the edema is due to elaboration of Vascular Endothelial Growth Factor (VEGF), a potent mediator of cerebro-vascular permeability (31). The receptors are over-expressed in glioblastomas. In metastasis, the accumulation of too much water is due to a vasogenic edema (31,32).
C=0.0076 x*T2 + D
Y=0.0076 x* T2 +68.481-------(2)
Watery component or Hydration portion “K” is the ratio of bound water (tied with the tissue) part to solid tissue part and the magnetic relaxation properties of bound and free part of water. C and D are two designated constants influenced by the hydration portion or component K. K value for white matter was measured as 0.37 at 37o C (30). In the T2 map, T2 value of the lesions within the central solid portion as well as 5mm and 10 mm away from center in the periphery were noted and tabulated.( Table 3.2 ).
As mentioned earlier the long T2 value of tissue is responsible by the “ free part” or “free phase” of water,. Out of 100 % of water roughly 20% of water of the total tissue is MRI invisible and motion restricted or” bound” (30,31).
QUANTIFICATION OF WATER IN THE PERIPHERAL PART OF LESIONS AS PERILESIONAL EDEMA
Relationship between T2 values of solid fraction of tissue and perilesional edema were tabulated in Table 3.2. The water content of glioma in the periphery (10 mm from the solid portion) as perilesional edema was found to be 14% more than the normal white matter water content. In glioblastomas it is about 18% and in metastasis it was almost 24% more than the normal white matter water content. In both glioblastoma or other malignant tumor the edema is due to elaboration of Vascular Endothelial Growth Factor (VEGF), a potent mediator of cerebro-vascular permeability. The receptors are over-expressed in aglioblastoma. In metastasis the accumulation of out of proportion water is due to the vasogenic edema (5,32). Fluid collection around the benign tumor or cysticercosis and tuberculoma is due to vasogenic edema.
Table 3.3 depicts high T2 values of CSF and edema fluid as they have high velocity protons. Due to this, the fluid takes a long time to hand over its energy (h) to the neighboring tissue structures (lattice) to complete the exchange of energy and the relaxation process is prolonged. Due to the sluggish velocity of the solid fraction of tissue, energy exchange is relatively quicker and therefore, solid tissues have a low T2 value.
CALCULATION OF ADC FROM ADC MAP
ADC values were calculated from the ADC map (taking b=0 and /or 1000 s/mm2) depending on the diffusion weighted image (10,11) and displayed in the scanner (Figure 3.3). ADC values were tabulated in Table 3.4.
APPARENT DIFFUSION COEFFICIENT (ADC VALUE)
METABOLITES OF MR SPECTROSCOPY
Choline-creatine ratio and creatine-NAA, lipid/lactate ratio, different metabolites like choline, NAA, creatine, MI of different patients have been quantified and put into a table (Table 3.5).
3.5. P C A
A two-dimensional analysis of the various tissues and diseases by principal components is helpful in discriminating the lesions (17). Three dimensional clustering (Figure 3.6) on the other hand by PCA is not always practical. There is overlapping of the values of diseases such as MS over the normal and gray and white matters which is also discrete in nature.
CLUSTERING OF THE PARAMETER AND DISEASES
In Figure 3.6 A and B clustering of the various diseases and tissues and tumors were illustrated (16) with the help of a special statistical program X L Stat.
RELATIONSHIP OF T2 VALUE AND PATHOLOGICAL LESION
In Table 4.1 T2 value and corresponding pathological lesions are described. Long T2 values are mainly seen in the CSF perilesional edema and plaques of multiple sclerosis.
The T2 values 5 mm within the lesions are depicted In Table 4.1. CSF, edema, arachnoid cysts and plaques of MS have a long T2 value. Gray matter and white matter show a range of T2 value as well. Glioma and glioblastoma show a low T2 value ranging from 90 ms to 123 ms. They show increased T2 value in the periphery 10 mm away from the solid part of the lesion. In metastasis the solid part of the lesion shows T2 values ranging from 147 to 150 ms. In the periphery long T2 values are noticed due to collection of water /fluid as edema.
We mentioned earlier that the water content of gliomas in the periphery (10 mm away from the solid component) in the perilesional edema was found to be 14% more than the normal white matter water (70 w/w) content. In glioblastomas it was found to be 18 % higher and in metastasis, it was almost 24% higher. In both glioblastoma or other malignant tumors the edema is believed due to elaboration of “Vascular Endothelial Growth Factor (VEGF)”, a strong mediator of cerebro-vascular permeability due to the over expression of the receptors in glioblastomas. In metastasis accumulation of out of proportion of water is due to a vasogenic edema. This is the only discriminating point between glioblastomas and metastasis.
As stated earlier a quantification relationship between the peri-lesional edema i.e. accumulation of fraction of water and T2 value of the lesion was derived:
Y= 0.0765X + 68.481 (2)
Where X signifies T2 value of the lesion and Y denotes fraction of accumulated water.
Usually in the central and part of the lesions like glioblastoma and metastasis T2 values remain low due to increased solid component and low water content. Diminished T2 values noted in astrocytoma III and IV (glioblastomas) and metastasis as well (34) are due to increased moveable lipids arising from the tissue necrosis and membrane breakdown. Low oxygen tension causes high lactate concentration as well and tissue disintegration.
CLINICAL APPLICATION OF T2 VALUE WITH COLOR CODED MAP IN T2 WEIGHTED INPUT IMAGES
Diagnosis or discrimination can be plausible from the T2 weighted ground truth images and T2 value can be interpreted from the T2 map.
T2 AND COLOR CODED MAPPING: DICOM images show no DATA loss compared to images saved in JPG, TIFF, BMP extension . Appropriate pixel gray shade value (GSV) can be noted if the image were saved in DICOM.
In the DICOM editor (Sante DICOM Editor3, Greece) input T2 weighted image was loaded for determination of GSV and color coded T2 mapping. It is necessary in certain images to run the color-coded map standardization of the images because of the non-uniformity of gray scale setting of the signal in some MRI machines, where pixel Gray Scale Value (GSV) of CSF of ventricles does not match with the 250 to 255 range. Figure 4.2 discerns the pixel GSV. Table 4.2 shows the range of pixel GSV and corresponding probable diseases or suspected pathological etiology.
After loading the input image a palette of the T2 color map can be displayed in the DICOM editor for color coded mapping (Figure4.3 and 4.4).
DISCRIMINATION OF MASSES OF MALIGNANT ORIGIN
The T2 weighted input ground truth image was loaded in a DICOM editor for color coded mapping. Red and bright red areas indicate malignant tissue (T2 value of 108 to 118 ms) of the mass and bluish white area indicates perilesional edema (T2 value of 210 to 241 ms). A T2 map derived from the MR machine is required to note the T2 value of the suspicious portion of the tissue and perilesional edema (Figure 4.3 and 4.4).
The color coded T2 map shows nodular and uneven wall of the metastasis from a lung cancer (Figure 4.5) or from a breast cancer demonstrating brilliant ruby color to light ruby color as an artificial shade superimposed on the T2 weighted image. T2 value is about 223 ms. Original MRS shows high lipid, lactate and choline peaks and almost no NAA.
The metastatic nodule showed a GSV of 196 to 210, determined in the DICOM editor and a T2 value around 223 to 227 ms and that of the perilesional edema (10 mm away from the nodule) was about 250 ms (4.5 D). This out of proportion water collection (25% more than the white matter water content) goes in favour of the diagnosis of metastasis. MRS shows increased choline, lipid and lactate peak and diminished peak of NAA.
EXAMINATION OF NON-MALIGNANT CASES DEMARCATION BETWEEN EPIDERMOID AND ARACHNOID CYST
Arachnoid cyst and epidermoid tumor can be differentiated by diffusion weighted image by showing restriction of diffusion. Demarcation can also be done from the T2 map. T2 values in arachnoid cysts vary from 175 to 275 ms whereas in epidermoid cysts it varies from 75 to 202 ms. T2 weighted image of an arachnoid cyst (Figure4.6A, B) and an epidermoid cyst (Figure4.6 C, D) were also examined with color coded T2 mapping (Figure 4.6 E and F). Arachnoid cysts has homogeneous internal fluid due to CSF whereas the internal fluid of epidermoid cyst looks variegated.
BRAIN ABSCESS AND MULTIPLE CYSTICERCOSIS
Diagnostic problem is encountered in nuro-cysticercosis and solid brain abscess/cysts (Figure 4.7 A, B. and C, D respectively). T2 color coded map generated may depict blue colored intra cystic fluid of high T2 value. Fluid of neuro-cysticersosis and solid abscess depict a T2 value of 280 to 290 ms. MR spectroscopy demonstrates high lipid and lactate peaks in the abscess.
The pyogenic membranes of an abscess and wall of the cysts can be demonstrated after color-encrypted mapping they are usually very smooth and thin (less than a millimeter). In metastasis the wall may be wavy, beaded uneven and thick.(Fig.4.7C, D).
OTHER BENIGN LESIONS (MULTIPLE SCLEROSIS)
Sometimes tumefactive MS, a demyelinating white matter chronic disease with tumor like plaques and demyelination creates a diagnostic problem (Figure 4.8 A, B, C and D). Spectroscopy also shows high choline and low NAA peak mimicking malignancy. In acute stage high lipid and lactate peaks are observed and are not observed in chronic and treated (remittance) patients (31). A T2 color coded image depicts a bluish white mass like plaque (Figure 4.8) with T2 values of 250 to 270 ms indicating presence of fluid. Multi voxel MR spectroscopy shows a tall choline peak and altered ratio of choline and creatine. Ratio of Choline and NAA along with lactate and lipid peaks remain unaltered.
DATA ANALYTICAL TECHNIQUE FOR LIVE PREDICTION:
i) If the supporting data are available then live prediction of diseases or of the tissue can be predicted in 90 to 95% by statistical methods eg ANN.
ii) Clustering of the diseases and tissues can also be plausible by Principal Component Analysis (PCA) or by K mean clustering.
PREDICTION OF DISEASES BY ANN:
Neural Tool (7.5 Palisade INC) is a Probabilistic Artificial Neural Network program (PNN) of non linear in nature and has been applied for live prediction purposes. ANN was introduced to evaluate different metabolic fractions, facts retrieved from MRI, and their ratio of MR spectroscopy, associated ADC values,T2 value of central solid and peripheral (Table 4.3) regions to virtually predict the pathological diagnosis(18,35).
ANN has exceptional and special data handling and learning capacity. All these qualities of ANN were exploited to predict the disease (24). Ten independent numeric variables were selected as inputs of ANN to create one dependent numerical variable or output (19). 10 nodes produced a single hidden layer (18,19) with 8 output nodes as target needed for prediction of gray and white matter, CSF, glioma, metastasis and cysts. (Figure 4.9).
INDEPENDENT VARIABLES AS INPUTS:
T2 Values within 5 mm of the lesion
T2 Values 10 mm away from the solid lesion
Quantification of metabolites (choline creatine, MI, NAA, lipid/ lactate)
Ratio of choline and NAA
Ratio of creatine and NAA
TO LIVE PREDICT (OR DECISION MAKING)
Type of Tissues, CSF and diseases as DEPENDENT VARIABLES
In the excel data sheet the dependent variables (disease) values are kept in the extreme left column (Table 4.3) and values of independent numeric variables (Usually RI value, choline NAA ratio or ADC value) in the right column.
A DATA SET MANAGER is prepared from the values of the excel sheet.
1. Values of table 5.3 to be trained and tested (statistics in Table 4.3).
2. Several attempts of prediction were done of the trained data set as testing.
3. If the T2 values were kept on the extreme right of the table as independent numerical variable the prediction of tissue and diseases was 100% (Table 4.4).
4. Prediction was accurate by 25% to 65% if the ratio of choline and creatine or ADC values was selected (Table 4.3) in the extreme right side of the data set.
In table 4.5 the number of incorrect prediction, classification rate, sensitivity and specificity has been tabulated using the 10- fold cross validation method. The classification rate observed using this method is quite high and very few errors have been observed in the prediction of test samples. Corresponding sensitivity and specificity have also been shown in the table. In most of the cases the sensitivity is a bit lower than the specificity. However, in a few exceptional cases the sensitivity has reached 100% where all the disease samples were identified. Since the grouping of the data was done in a random fashion, all groups do not have equal number of sick patients and in a few samples the number of disease samples were lower than the rest. Hence, it can be concluded that the types of disease’ do depend on T2 values of the image more than the ADC values and other quantified metabolic products or neuro transmitter like creatine, lipid and lactate, NAA and choline,.
From the Figure 4.10 it is observed that the mean square error of the data during training decreases with iteration and finally becomes constant. The error is very small between 0.15 to 0.2 units. Thus, the dataset has been trained such that the prediction error reaches a minimum value and then testing has been conducted using this trained model.
CLUSTERING OF THE DATA BY PCA
The PCA plot has been implemented on the data set (Table 4.3). It has been observed that the set belonging to a particular tissue or disease formed well defined clusters. Since these data sets are well clustered, it can be stated that these variables can be used to distinguish the diseases from one another and may be used for prediction of diseases. The PCA plot has been shown in the Figure 4.11.
Unlike ANN, PCA is statistically not very competent to discriminate the tissue or diseases only depict them one-dimensionally.
An alternative noninvasive diagnostic procedure was described by mapping the T2 weighted MR images of the brain using various pathological lesions with false color-coded map discriminating the benign from the malignant diseases.
ANN and PCA have been implemented to live predict the tissue character and diseases utilizing the data of ground truth input images. p-values (Pearson Phi) are shown in Table 4.6 and Figure 4.12. For T2 values the p value is 0.002. It is maximum in relation to ADC values.
Some pitfalls are noticed between discrimination of metastasis and glioblastoma. These non-invasive procedures can be implemented as presumptive diagnosis before doing stereotactic biopsy. Thus risk and costs can be minimized.
We are thankful to the Palisade Inc for providing the Neural Tool version 7.5 free of cost for the research work.