Simulation Of Contrast Enhancement Of Various Brain Lesions (Without Iv Gadolinium) By Using The Neural Network
T Biswas. Simulation Of Contrast Enhancement Of Various Brain Lesions (Without Iv Gadolinium) By Using The Neural Network. The Internet Journal of Radiology. 2012 Volume 14 Number 1.
To improve the visibility of internal body structures, MRI contrast agents are used. Gadolinium (Gd3+), a paramagnetic substance, is the most commonly used compounds for contrast enhancement. Chemically Gd is a rare earth element (Z-64)(1-3) that is toxic in its free state, but when it is bound to DTPA by chelation the problem of toxicity is solved (3,4). This paramagnetic substance has small local magnetic fields which cause a shortening of the relaxation times of the surrounding protons; this is known as proton relaxation enhancement effect (2). Gd-based MRI contrast agents alter the T1 relaxation time of tissues (spin-lattice relaxation time of protons) located nearby and in body cavities where they are present(5-9). Paul Lauterbur and his associates(10) were the first to demonstrate the feasibility of using paramagnetic contrast agents to improve tissue discrimination in MRI. Depending on the image weighting, this can give a higher or lower signal of two tissues which can be better differentiated(11). Most clinically used MRI contrast agents work through shortening the relaxation time. The T1 shortening is due to an increase in rate of stimulated emission from high energy states (aligned anti-parallel to the main field) to low energy states (parallel to the main field). The source of the stimulation is thermal vibration from the strongly magnetic metal ions, which create oscillating electromagnetic fields at frequencies corresponding to the energy differences between the spin states (via E = h?)(3,7,11).
MRI contrast agents are administered by injection into the blood stream for brain tumor and lesion enhancement associated with the degradation of the blood=brain barrier. Due to their hydrophilic character, gadolinium chelates do not pass the blood-brain barrier(5,8). Thus these are useful in enhancing lesions and tumors where the Gd leaks out (Figures 1 and 2). In the rest of the body, the Gd initially remains in the circulation before being distributed into the interstitial space or eliminated by the kidney(12).
Gadolinium chelates are extremely well tolerated by the vast majority of patients in whom they are injected. Acute, adverse reactions are encountered with a lower frequency than is observed after administration of iodinated contrast media (7). A serious complication like nephrogenic systemic fibrosis (NSF) and fibrosis in various tissues and organs in the body can occur(13-15). Contrast-mediated nephrotoxicity (CMN) is another complication. The frequency of all acute, adverse events after an injection of 0.1 or 0.2 mmol/kg of gadolinium chelate ranges from 0.07% to 2.4%(15). The vast majority of these reactions are mild. Severe, life-threatening anaphylactoid or non-allergic anaphylactic reactions are exceedingly rare (0.001% to 0.01%)(16).
Assessment of contrast enhancement on T1 weighted images of normal and various brain lesions was the essential element of our study (12). Statistical analysis of the pre- and post- contrast brain lesions, the intensity of signal in gray shades and the extent of contrast enhancement in tissue depends on the gadolinium content within each voxel(7). We measured the magnitude of longitudinal relaxation value (T1 relaxation time) in gray and white matters and in brain tumors before and after administration of a contrast agent using echo-planar inversion recovery sequences(17-19). Echo-planar imaging, a fast imaging-acquisition method, provides inversion recovery images more rapidly. However it could be useful to know the scale of T1 shortening. Inversion recovery (IR) sequences are sensitive to T1 contrast and the most precise means of measuring T1 relaxation rates in tissue in vivo.
A multi-slice EPI IR sequence was used to assess T1 relaxation times of various brain lesions and brain tumors before and after administration of Gd.
Our purpose was to assess whether there was any statistical relationship in the magnitude of changes in the T1 relaxation of brain tumors before and after the contrast material was administered, and whether enhancement-like effects of the lesions could be reproduced mathematically with the help of neural networks and without the IV contrast injection.
T1 relaxation and signal intensity, as represented by the corresponding gray shade values (out of 256 shades), of various brain lesions before and after IV contrast were tabulated. Using independent component analysis (ICA), a statistical technique, the complex dataset was decomposed into independent sub-parts.
With the help of K-mean clustering, the data were classified. A mapping function was generated that corresponded between these independent components and the cross-sectional data by using the neural network after training it with sample datasets (11,12). The training samples for the network were selected using K–mean clustering. A mapping of the lesion was overlaid on the T1 weighted image.
Materials And Methods
Brain T1 mapping is important to determine the T1 relaxation value of brain pathology. We used conventional T1 mapping techniques based on inversion recovery spin echo (IRSE) and its counterpart, inversion recovery spin echo (EPI IR) imaging. Both of them produced good quality images. Multiple inversion times were applied to accurately estimate the T1 relaxation value of white and gray matter and various brain lesions and tumors. All images were produced in a 3T Siemens Magnetom and a 0.3T AIRISII Magnet, (Hitachi, Japan). Conventional axial T1-weighted localizing image, conventional axial T1, proton density and T2 weighted images were obtained. Six consecutive slices were selected for EPI IR with 5 mm thickness without a gap, TI (time to inversion recovery) 50-1500ms with a step of 150ms, TR 15000ms, and TE of 30ms with a partial k-space acquisition. A matrix size of 256X256 and FOV 25X50cm was used. In the 0.3T magnet a matrix size of 256X128, 5mm slice thickness with 5 mm gap and 30 cm field of view (FOV) were used.
The IRSE protocol had TE=15ms, TR-2500 ms, receiver bandwidth =
SI=a (1-2e (-TI/T1))-b). Equation 1
We examined 21 patients after getting proper institutional ethics and consents from the patients (Table 1).
Before giving contrast we routinely produced T1 maps by conventional IRSE and EPI IR of the entire brain and determined the T1 value of gray/white matter, various part of the lesions and brain tumor. After that each patient received 0.1mmol/Kg of GD DTPA (Magnevist –Schering) by a Power Injector (Magmedix) at a rate of 5ml/second and was imaged by conventional IRSE to produce the T1 map.
Regions of interest presenting enhancing and/or non-enhancing parts of tumors were selected by viewing the post-contrast conventional SE images. Regions of interest containing white and gray matter areas were identified in the contralateral hemisphere. T1 values of the various part of the tumor were directly measured from the T1 map produced by IRSE and EPI IR protocols and tabulated (Table 2).
A segmentation technique was applied to decompose the image into mosaic form of the gray shades of the pre- and post-contrast parts of the normal brain tissue (such as gray and white matter) and brain tumors or lesions (Figure 3 A and B).
By applying K-mean clustering the gray shade values (out of 256) of various parts were determined and tabulated. K-mean clustering classified the data based on attributes/features. Each attribute represented the T1 relaxation value of tissue before and after contrast and gray shade value (out of 256 shades) as the signal. The grouping is done by minimizing the sum of squares of distances between data and the corresponding cluster centroid. Then pre- and post-contrast values of the tissues were transferred into the Excel sheet of the neural network program.
Neural Network Design and Steps
We followed the standard steps for designing neural networks to predict the signal and gray shade values of unknown tissues in four application areas: function fitting, pattern recognition, clustering, and time series analysis. The work flow for any of these problems had seven primary steps:
To collect data from T1 mapping and k mean clustering.
To create the network
To configure the network
To initialize the weights and biases
To train the network
To test and Validate the network
To use the network for prediction of value of unknown tissue.
A data set manager is prepared in Microsoft Excel: neural networks that contain 5 variables, 32 cells per variable containing known data. A trial version of the neural network program Neural Tool version 5.7 of the Palisade Inc(20) was used which imitated the brain-like functions in order to learn the structure of the data provided.
Out of 5 variables
A. One independent category variable (of tissue),
B. Three independent numeric category variables such as T1 value (in ms) of
i) Pre-contrast ii) post-contrast Tissue iii) gray shade signal value of pre- and post-contrast tissue (out of 256 shades)
C. One dependent numeric category variable, which ultimately produced live predict variables of post contrast values.
After understanding the data, the program worked through the following steps: analysis of the data, training the network on the data provided, testing the network for accuracy, and making intelligent predictions of the signal (gray values of the pixel out of 256 shades) from the new input of the T1 value of the unknown (target tissue).
Contrast Enhancement Simulation
A radial basis function network was then used to generate a contrast enhanced map overlay on the background T1 weighted image. It produced correspondences between independent components of gray shades of the pre-contrast T1 weighted image and live prediction of the dependent variable of gray shades (out of 256) of tumor/lesion of the T1 weighted image, pixel by pixel. The training samples for the neural network were selected using the k-mean clustering algorithm to give a contrast-like effect of the image of the tumor.
The mean T1 relaxation time in tumors was 1416 ms before administration of contrast, and 631 ms after the injection compared to 646 ms in white matter. T1 relaxation time decreased from 1179 ± 130 ms to 926 ± 126 ms in gray matter (p<0.001), and from 706 ± 107 ms to 646 ± 27 ms in white matter (p<0.001) after the 0.1 mmol/kg of contrast. Post-contrast T1 relaxation times in tumors showed considerable variation and remained, on average, relatively long compared to white matter.
In figure 4A, the relationship of pre- and post-contrast T1 values of pathological tissue was established:
Post-contrast T1 relaxation = - 0.514 x T1 Pre-contrast relaxation + 1427.6 (Equation 2)
Figure 4B depicts the relationship of the pre- and post-contrast signals expressed in gray scales:
Post-contrast T1 signal = 0.0068x2 - 2.2861xPrecontrast T1 signal + 392.18 (Equation 3)
In Table 2, pre-contrast T1, post-contrast T1, pre- and post-contrast signal in gray scale are tabulated.
Analysis Of T1 Relaxation
T1 relaxation times of the various parts of tumors, contralateral gray and white matters were obtained from the T1 map (21,22). Considerable variation in T1 relaxation times in tumors in different patients was noticed before the contrast injection (22). Detectable tumor enhancement was noticed in all 21 patients in conventional T1 weighted images. After the contrast injection, T1 relaxation time shortened both in the tumors and in the contralateral gray and white matters. From Table 2, it is evident that T1 relaxation in the tumor is more marked compared to gray/white matter, possibly due to intratumoral, vasogenic edema and perilesional edema which interfere with the heat dissipation during relaxation. After contrast, shortening of T1 relaxation time was noted in most of the tumors (17). Analyzing the percentage of T1 relaxation after contrast enhancement, it is evident that it is at a minimum in white matter (0.15 to 12.84%, mean-3.3%) and maximum (79%, mean) in case of tumors(23).
We used the General Regression Neural Networks (GRNN) numeric predictor configuration.
Architecture of a PNN/GRNN Network
In 1990, Donald F. Specht proposed a method to formulate the weighted-neighbor method described above in the form of a neural network(24-27). He called this a “Probabilistic Neural Network.” Here is a diagram of a PNN/GRNN network (Figure 6)(28,29):
All PNN/GRNN networks have four layers: input, hidden, pattern/summation and decision/prediction.
Input layer — Pre- and post-contrast T1 values of gray and white matters, various tumors and brain lesions.
The input neurons/nodes (or processing before the input layer) standardize the range of the values by subtracting the median and dividing by the interquartile range used to summarize the extent of the spread of the data (variability between 75 th and 25 th percentiles). The input neurons then feed the values to each of the neurons in the hidden layer(30). In table 3, a summary of the neural network is provided.
Table 4 depicts how the dataset manager trains the inputs and tests to predict. Most of the predictions were found to be good. Residual values (between original and predicted value) show a range from -0.97 to -19.28.
Hidden layer — There were 26 cases and, during training, 98 trials were applied. During testing, the tool tried 6 predictions, out of which a 16.66% were found bad after analyzing the residual values (actual values minus predicted values). Testing data generated by the dataset manager along with histogram bins are depicted in Table 5.
The trained Neural Net’s prediction (auto-testing) of unknown post-contrast T1 values of pathological tissues when pre-contrast T1 values and signal shades were provided are in Table 6A and 6B .
In Table 6C are the values for R-square (training) and root mean square error (training and testing) of Linear Predictor and Neural Net
6C.R-Square (Training), Root mean Sq error(training), Root mean Sq (Testing)
Figures 7A, 7B and 7C discern the relationship of Predicted versus Actual, Residual versus Actual (Training) and Predicted versus Actual testing.
Simulation Of Contrast Enhancement
Figures 8A, 8B, and 8C show the non-contrast T1 weighted, contrasted (post-GD) image and brain mapping of contrast-like effect. Contrast simulation and original GD enhancement have similarities in shape and pattern, as depicted in 8B and 8C respectively.
Figure 9C depicts the simulation of contrast enhancement of a glioblastoma overlaying the T1 weighted plain image
Figure 10 shows a contrast enhancement effect in: A. an intra-ventricular meningioma B. a medulloblastoma C. an oligodendroglioma and D. Arachnoid cyst respectively. Increased enhancement of the gyri is also noted. A contrasted image of lymphoma (Figure 11B)overlaid on a T1 pre-contrast image (11B).
DICOM FORMAT VS JPEG FORMAT
Smooth and regular enhancement is noted in the peripheral component of the tumor (9A) in
GD enhancement whereas the contrast-like effect, though very similar, shows an irregular outline and is interrupted in texture (9B). This can be explained by the fact that the MR images are displayed in Digital Imaging and Communications in Medicine (DICOM) format compared to exported images in JPEG (JPEG baseline (lossy 8 bits/pixel) compression) format used for simulation effect. The DICOM image is a 16-bit gray scale image (signed/unsigned). This means each pixel is represented by 16 bits, 216 or 65,536 shades/pixel in DICOM compared to 8-bit or 28 = 256 shades/pixel of gray level combinations available in a JPEG saved image, so a smooth and regular outline is not depicted in the simulation effect, unlike in the GD-enhanced image. The problem is extracting raw data from the DICOM image available for the simulation effect.
An attempt was successfully made to get an enhancement effect in various brain lesions statistically, utilizing a neural network and without IV contrast. I am working to retrieve the raw data and the signal from the DICOM format directly so that an accurate, contrast-like effect can be generated.
I am thankful to Palisade Inc, Asia- Pacific Pty Inc, Sydney, NSW 2000 Australia, for providing a 15 days Trial version of Neural Tool (5.7).I sincerely thank MR Samir Roy of Biswas Scan Centre, DR Sadallah Ramadan of New Castle University, Australia.