Blended-Transfer Learning for Compressed-Sensing Cardiac CINE MRI

Purpose@#To overcome the difficulty in building a large data set with a high-quality in medical imaging, a concept of 'blended-transfer learning' (BTL) using a combination of both source data and target data is proposed for the target task. @*Materials and Methods@#Source and target tasks were defined as training of the source and target networks to reconstruct cardiac CINE images from undersampled data, respectively. In transfer learning (TL), the entire neural network (NN) or some parts of the NN after conducting a source task using an open data set was adopted in the target network as the initial network to improve the learning speed and the performance of the target task. Using BTL, an NN effectively learned the target data while preserving knowledge from the source data to the maximum extent possible. The ratio of the source data to the target data was reduced stepwise from 1 in the initial stage to 0 in the final stage. @*Results@#NN that performed BTL showed an improved performance compared to those that performed TL or standalone learning (SL). Generalization of NN was also better achieved. The learning curve was evaluated using normalized mean square error (NMSE) of reconstructed images for both target data and source data. BTL reduced the learning time by 1.25 to 100 times and provided better image quality. Its NMSE was 3% to 8% lower than with SL. @*Conclusion@#The NN that performed the proposed BTL showed the best performance in terms of learning speed and learning curve. It also showed the highest reconstructed-image quality with the lowest NMSE for the test data set. Thus, BTL is an effective way of learning for NNs in the medical-imaging domain where both quality and quantity of data are always limited.

Blended-Transfer Learning for Compressed-Sensing Cardiac CINE MRI

Purpose@#To overcome the difficulty in building a large data set with a high-quality in medical imaging, a concept of 'blended-transfer learning' (BTL) using a combination of both source data and target data is proposed for the target task. @*Materials and Methods@#Source and target tasks were defined as training of the source and target networks to reconstruct cardiac CINE images from undersampled data, respectively. In transfer learning (TL), the entire neural network (NN) or some parts of the NN after conducting a source task using an open data set was adopted in the target network as the initial network to improve the learning speed and the performance of the target task. Using BTL, an NN effectively learned the target data while preserving knowledge from the source data to the maximum extent possible. The ratio of the source data to the target data was reduced stepwise from 1 in the initial stage to 0 in the final stage. @*Results@#NN that performed BTL showed an improved performance compared to those that performed TL or standalone learning (SL). Generalization of NN was also better achieved. The learning curve was evaluated using normalized mean square error (NMSE) of reconstructed images for both target data and source data. BTL reduced the learning time by 1.25 to 100 times and provided better image quality. Its NMSE was 3% to 8% lower than with SL. @*Conclusion@#The NN that performed the proposed BTL showed the best performance in terms of learning speed and learning curve. It also showed the highest reconstructed-image quality with the lowest NMSE for the test data set. Thus, BTL is an effective way of learning for NNs in the medical-imaging domain where both quality and quantity of data are always limited.

Biases in the Assessment of Left Ventricular Function by Compressed Sensing Cardiovascular Cine MRI

PURPOSE: We investigate biases in the assessments of left ventricular function (LVF), by compressed sensing (CS)-cine magnetic resonance imaging (MRI). MATERIALS AND METHODS: Cardiovascular cine images with short axis view, were obtained for 8 volunteers without CS. LVFs were assessed with subsampled data, with compression factors (CF) of 2, 3, 4, and 8. A semi-automatic segmentation program was used, for the assessment. The assessments by 3 CS methods (ITSC, FOCUSS, and view sharing (VS)), were compared to those without CS. Bland-Altman analysis and paired t-test were used, for comparison. In addition, real-time CS-cine imaging was also performed, with CF of 2, 3, 4, and 8 for the same volunteers. Assessments of LVF were similarly made, for CS data. A fixed compensation technique is suggested, to reduce the bias. RESULTS: The assessment of LVF by CS-cine, includes bias and random noise. Bias appeared much larger than random noise. Median of end-diastolic volume (EDV) with CS-cine (ITSC or FOCUSS) appeared −1.4% to −7.1% smaller, compared to that of standard cine, depending on CF from (2 to 8). End-systolic volume (ESV) appeared +1.6% to +14.3% larger, stroke volume (SV), −2.4% to −16.4% smaller, and ejection fraction (EF), −1.1% to −9.2% smaller, with P < 0.05. Bias was reduced from −5.6% to −1.8% for EF, by compensation applied to real-time CS-cine (CF = 8). CONCLUSION: Loss of temporal resolution by adopting missing data from nearby cardiac frames, causes an underestimation for EDV, and an overestimation for ESV, resulting in underestimations for SV and EF. The bias is not random. Thus it should be removed or reduced for better diagnosis. A fixed compensation is suggested, to reduce bias in the assessment of LVF.

Simultaneous Unwrapping Phase and Error Recovery from Inhomogeneity (SUPER) for Quantitative Susceptibility Mapping of the Human Brain

PURPOSE: The effect of global inhomogeneity on quantitative susceptibility mapping (QSM) was investigated. A technique referred to as Simultaneous Unwrapping Phase with Error Recovery from inhomogeneity (SUPER) is suggested as a preprocessing to QSM to remove global field inhomogeneity-induced phase by polynomial fitting. MATERIALS AND METHODS: The effect of global inhomogeneity on QSM was investigated by numerical simulations. Three types of global inhomogeneity were added to the tissue susceptibility phase, and the root mean square error (RMSE) in the susceptibility map was evaluated. In-vivo QSM imaging with volunteers was carried out for 3.0T and 7.0T MRI systems to demonstrate the efficacy of the proposed method. RESULTS: The SUPER technique removed harmonic and non-harmonic global phases. Previously only the harmonic phase was removed by the background phase removal method. The global phase contained a non-harmonic phase due to various experimental and physiological causes, which degraded a susceptibility map. The RMSE in the susceptibility map increased under the influence of global inhomogeneity; while the error was consistent, irrespective of the global inhomogeneity, if the inhomogeneity was corrected by the SUPER technique. In-vivo QSM imaging with volunteers at 3.0T and 7.0T MRI systems showed better definition in small vascular structures and reduced fluctuation and non-uniformity in the frontal lobes, where field inhomogeneity was more severe. CONCLUSION: Correcting global inhomogeneity using the SUPER technique is an effective way to obtain an accurate susceptibility map on QSM method. Since the susceptibility variations are small quantities in the brain tissue, correction of the inhomogeneity is an essential element for obtaining an accurate QSM.

Fast Cardiac CINE MRI by Iterative Truncation of Small Transformed Coefficients

PURPOSE: A new compressed sensing technique by iterative truncation of small transformed coefficients (ITSC) is proposed for fast cardiac CINE MRI. MATERIALS AND METHODS: The proposed reconstruction is composed of two processes: truncation of the small transformed coefficients in the r-f domain, and restoration of the measured data in the k-t domain. The two processes are sequentially applied iteratively until the reconstructed images converge, with the assumption that the cardiac CINE images are inherently sparse in the r-f domain. A novel sampling strategy to reduce the normalized mean square error of the reconstructed images is proposed. RESULTS: The technique shows the least normalized mean square error among the four methods under comparison (zero filling, view sharing, k-t FOCUSS, and ITSC). Application of ITSC for multi-slice cardiac CINE imaging was tested with the number of slices of 2 to 8 in a single breath-hold, to demonstrate the clinical usefulness of the technique. CONCLUSIONS: Reconstructed images with the compression factors of 3-4 appear very close to the images without compression. Furthermore the proposed algorithm is computationally efficient and is stable without using matrix inversion during the reconstruction.

Cardiac Magnetic Resonance Imaging Using Multi-physiological Intelligent Trigger System

PURPOSE: We proposed a multi-physiological signals based real-time intelligent triggering system(MITS) for Cardiac MRI. Induced noise of the system was analyzed. MATERIALS AND METHODS: MITS makes cardiac MR imaging sequence synchronize to the cardiac motion using ECG, respiratory signal and second order derivative of SPO2 signal. Abnormal peaks due to arrhythmia or subject's motion are rejected using the average R-R intervals and R-peak values. Induced eddy currents by gradients switching in cardiac MR imaging are analyzed. The induced eddy currents were removed by hardware and software filters. RESULTS: Cardiac MR images that synchronized to the cardiac and respiratory motion are acquired using MITS successfully without artifacts caused by induced eddy currents of gradient switching or subject's motion or arrhythmia. We showed that the second order derivative of the SPO2 signal can be used as a complement to the ECG signals. CONCLUSION: The proposed system performs cardiac and respiratory gating with multi-physiological signals in real time. During the cardiac gating, induced noise caused by eddy currents is removed. False triggers due to subject's motion or arrhythmia are rejected. The cardiac MR imaging with free breathing is obtained using MITS.

Water-Fat Imaging with Automatic Field Inhomogeneity Correction Using Joint Phase Magnitude Density Function at Low Field MRI

PURPOSE: A new inhomogeneity correction method based on two-point Dixon sequence is proposed to obtain water and fat images at 0.35T, low field magnetic resonance imaging (MRI) system. MATERIALS AND METHODS: Joint phase-magnitude density function (JPMF) is obtained from the in-phase and out-of-phase images by the two-point Dixon method. The range of the water signal is adjusted from the JPMF, and 3D inhomogeneity map is obtained from the phase of corresponding water volume. The 3D inhomogeneity map is used to correct the inhomogeneity field iteratively. RESULTS: The proposed water-fat imaging method was successfully applied to various organs. The proposed 3D inhomogeneity correction algorithm provides good performances in overall multi-slice images. CONCLUSION: The proposed water-fat separation method using JPMF is robust to field inhomogeneity. Three dimensional inhomogeneity map and the iterative inhomogeneity correction algorithm improve water and fat imaging substantially.

Study of Motion Effects in Cartesian and Spiral Parallel MRI Using Computer Simulation

PURPOSE: Motion effects in parallel magnetic resonance imaging (MRI) are investigated. Parallel MRI is known to be robust to motion due to its reduced acquisition time. However, if there are some involuntary motions such as heart or respiratory motions involved during the acquisition of the parallel MRI, motion artifacts would be even worse than those in conventional (non-parallel) MRI. In this paper, we defined several types of motions, and their effects in parallel MRI are investigated in comparisons with conventional MRI. MATERIALS AND METHODS: In order to investigate motion effects in parallel MRI, 5 types of motions are considered. Type-1 and 2 are periodic motions with different amplitudes and periods. Type-3 and 4 are segment-based linear motions, where they are stationary during the segment. Type-5 is a uniform random motion. For the simulation, Cartesian and spiral grid based parallel and non-parallel (conventional) MRI are used. RESULTS: Based on the motions defined, moving artifacts in the parallel and non-parallel MRI are investigated. From the simulation, non-parallel MRI shows smaller root mean square error (RMSE) values than the parallel MRI for the periodic (type-1 and 2) motions. Parallel MRI shows less motion artifacts for linear (type-3 and 4) motions where motions are reduced with shorter acquisition time. Similar motion artifacts are observed for the random motion (type-5). CONCLUSION: In this paper, we simulate the motion effects in parallel MRI. Parallel MRI is effective in the reduction of motion artifacts when motion is reduced by the shorter acquisition time. However, conventional MRI shows better image quality than the parallel MRI when fast periodic motions are involved.

Development of MRI Phantom for Assessing MR Image Quality

PURPOSE : To evaluate MR image qualities we developed a new MRI phantom with the fixation structures necessary to position it into coil firmly. MATERIALS AND METHODS : We designed MRI phantom for eight evaluation items such as slice thickness accuracy, high contrast spatial resolution, low contrast object detectability, geometry accuracy, slice position accuracy, image intensity uniformity, percent signal ghosting and signal to noise ratio. For the positioning of phantom at coils, the fixation structures were set up on the surface of phantom. Six different MRI units were used for test the possibility for the clinical application and their image qualities were evaluated. RESULTS : We acquired appropriate MR image qualities enough for the evaluation on all used MR units and confirmed that their evaluations were within reliable values compared to real ones for some items. The positioning of our phantom into head coils with fixation structures worked well for proper imaging. CONCLUSION : We found that our prototype of MRI phantom had the possibility of clinical application for MR image quality assessment.

High Resolution MR Images from 3T Active-Shield Whole-Body MRI System

PURPOSE: Within a clinically acceptable time frame, we obtained the high resolution MR images of the human brain, knee, foot and wrist from 3T whole-body MRI system which was equipped with the world first 3T active shield magnet. MATERIALS AND METHODS: Spin echo (SE) and Fast Spin Echo (FSE) images were obtained from the human brain, knee, foot and wrist of normal subjects using a homemade birdcage and transverse electromagnetic (TEM) resonators operating in quadrature and tuned to 128 MHz. For acquisition of MR images of knee, foot and wrist, we employed a homemade saddle shaped RF coil. Typical common acquisition parameters were as follows: matrix= 512x512, field of view (FOV) = 20 cm, slice thickness = 3 mm, number of excitations (NEX) = 1. For T1-weighted MR images, we used TR= 500 ms, TE = 10 or 17.4 ms. For T2-weighted MR images, we used TR=4000 ms, TE = 108 ms. RESULTS: Signal to noise ratio (SNR) of 3T system was measured 2.7 times greater than that of prevalent 1.5T system. MR images obtained from 3T system revealed numerous small venous structures throughout the image plane and provided reasonable delineation between gray and white matter. CONCLUSION: The present results demonstrate that the MR images from 3T system could provide better diagnostic quality of resolution and sensitivity than those of 1.5T system. The elevated SNR observed in the 3T high field magnetic resonance imaging can be utilized to acquire images with a level of resolution approaching the microscopic structural level under in vivo conditions. These images represent a significant advance in our ability to examine small anatomical features with noninvasive imaging methods.

Clinical Validity of the Domestic EEG and Mapping System(Neuronics) / 수면정신생리

The clinical validity of a korean EEG and EP mapping system(Neuronics) was evaluated with schizophrenic patients(n=20), normal controls(n=19), and 10 patients with central nervous system disease(8 patients with cerebrovascular accident, 1 patient with brain mass, and 1 patient with periodic paralysis). In the normal control group, the pattern of resting computerzied EEG with eyes closed showed normal parieto-occipital dominance of alpha wave. Compared with normal controls, schizophrenic patients had more delta activity in the frontal region, and less alpha activity especially in the parieto-occipital region. In most cases patients with cortical organic lesions(n=5) revealed increased delta and theta activity and decreased alpha activity on the lesion areas. These findings were compatible with their MRI and clinical findings. However in the cases of subcortical lesions(n=5) EEG showed various findings which suggest diverse influences of subcortical abnormalities on cortical activities. The P300 of schizophrenic group was smaller and more delayed than those of normal controls. These results are generally compatible with the previous studies using other EEG and EP mapping systems consequenty and suggest that the this EEG and EP mapping system(Neuronics) has clinical validity.

Magnetic Resonance Imaging in Neurologic Diseases

Magrjetic resonance (MR) imagings with 0.15 Tesla resistive magnet developed by Korea Advanced Institute of 5cience were performed in 27 patients with various neurologic diseases and compared with X-ray computed tomography (CT). The purpose óf the paper is to evaluate the image quality, the diagnostic value and limitation, and the optimal pulse sequence of MR imagings with a resistive magnet. The MR images were obtained by using a variety of pulse sequence with spin echo technique includïng saturation recovery, T2-weighted spin echo, and/or inversion recovery with various pulse repetition(TR) and echo delay (TE) times. The M R imaging demonstrated the capability of detecting the lesions shown on CT in all cases and also detected an additional finding in one case (multiple sclerosis) which was not seen on CT. The MR imaging appeared to be more useful than CT in the evaluation of syringomyelia of spinal cord and white matter disease, while it failed to demonstrate small calcific lesion or inflammatory nodule (less than 1 cm) shown on CT and has shown somewhat poor contrast resolution in the case of meningioma. The spatial resolution of saturation recovery images was similar or superior to CT, whereas the contrast resolution of saturation recovery was inferior to CT. While the saturation recovery images have shown false negative findings in 5 patients (19%), the inversion recovery and T2-weighted spin echo have shown consistently positive findings. The inversion recovery and T2-weighted spin echo images demonstrated better contrast discrimination between normal and pathologic conditions than the saturation recovery images, but somewhat poorer spatial resolution. Authors suggest that the MR images of both the saturation recovery with 300/30 and T2-weighted spin echo with 1000/90 be used as a routine procedure and additional iversion recovery of 1300/300/30 sequence as a option if white matter disease is suspected.