Taehoon Shin

While magnetic resonance imaging (MRI) is effective for detecting early-stage osteonecrosis of the femoral head (ONFH), it is often expensive and less accessible; conversely, radiography is more widely accessible but has limited sensitivity for early-stage diagnosis. We developed a deep learning approach using radiographic images to effectively classify ONFH stages, providing a more accessible method for early diagnosis and disease stage differentiation. The dataset consisted of 909 hip radiographs, yielding 1818 femoral head images (grade 0:1495; grade 1:80; grade 2:114; grade 3:93; grade 4:36). A U-Net model was used to segment the femoral heads, achieving a Dice similarity coefficient (DSC) of 0.977 on the test set, allowing precise localization of the region of interest. A variational autoencoder (VAE) was then trained using 1270 grade-0 images for training and 112 for validation to construct a normative latent distribution representing healthy femoral heads. When ONFH data from all grades were projected into the latent space, significant differences in Mahalanobis distance distributions were observed across most ONFH stages. No significant difference was found between grades 0 and 1 (p = 0.06), consistent with known radiographic subtlety. However, grades 2–4 showed significant deviation from grade 0, and significant differences were also observed among mid- and late-stage grades. These findings demonstrate that the proposed method effectively separates healthy and diseased femoral heads and captures gradewise structural progression within the latent space. This radiograph-based normative modeling approach offers an accessible alternative to MRI for ONFH stage differentiation, particularly in resource-limited clinical environments. Although early-stage differentiation remains challenging, the results highlight the potential of radiograph-based deep learning systems to improve diagnostic efficiency and support future automated ONFH staging workflows.

Purpose: White matter hyperintensities (WMHs) are key neuroimaging markers of cerebral small vessel disease (cSVD), associated with cognitive decline and increased stroke risk. We aimed to investigate whether carotid time-of-flight (TOF) magnetic resonance angiography (MRA), a routinely acquired and non-invasive vascular imaging modality, can be utilized to independently predict WMH burden using deep learning.Methods: We developed a deep learning-based framework to predict WMH presence and severity using only 3D carotid TOF MRA. Two classification tasks were defined: binary (grade 0 vs. grades 1–3) and three-class (grade 0, 1, 2–3) classification. Four model architectures— simple fully convolutional network (SFCN), ResNet10, MedicalNet, and Medical Slice Transformer—were evaluated. To enhance model interpretability, we performed saliency mapping and occlusion analysis.Results: SFCN performed the best, achieving an accuracy of 76.5% and an area under the receiver operating characteristic curve (AUC) of 0.874 in binary classification, along with a 63.5% accuracy and a 0.827 AUC in WMH severity classification. Interpretability analyses confirmed that models predominantly focused on carotid vessel regions, which supports known vascular associations with WMH burden.Conclusion: Carotid TOF MRA alone can serve as a predictive marker for WMH burden when analyzed using deep learning. This approach highlights the potential utility of extracranial carotid imaging as a non-invasive surrogate for early and accessible assessment of cerebrovascular risk.

Introduction: Aristotle illusion is a well-known tactile illusion which causes the perception of one object as two. EEG analysis was employed to investigate the neural correlates of Aristotle illusion, yet was limited due to low spatial resolution of EEG. This study aimed to identify brain regions involved in the Aristotle illusion using functional magnetic resonance imaging (fMRI) and deep learning-based analysis of fMRI data.Methods: While three types of tactile stimuli (Aristotle, Reverse, Asynchronous) were applied to thirty participants’ fingers, we collected fMRI data, and recorded the number of stimuli each participant perceived. Four convolutional neural network (CNN) models were trained for perception-based classification tasks (the occurrence of Aristotle illusion vs. Reverse illusion, the occurrence vs. absence of Reverse illusion), and stimulus-based classification tasks (Aristotle vs. Reverse, Reverse vs. Asynchronous, and Aristotle vs. Asynchronous).Results: Simple fully convolution network (SFCN) achieved the highest classification accuracy of 68.4% for the occurrence of Aristotle illusion vs. Reverse illusion, and 80.1% for the occurrence vs. absence of Reverse illusion. For stimulus-based classification tasks, all CNN models yielded accuracies around 50% failing to distinguish among the three types of applied stimuli. Gradient-weighted class activation mapping (Grad-CAM) analysis revealed salient brain regions-of-interest (ROIs) for the perception-based classification tasks, including the somatosensory cortex and parietal regions.Discussion: Our findings demonstrate that perception-driven neural responses are classifiable using fMRI-based CNN models. Saliency analysis of the trained CNNs reveals the involvement of the somatosensory cortex and parietal regions in making classification decisions, consistent with previous research. Other salient ROIs include orbitofrontal cortex, middle temporal pole, supplementary motor area, and middle cingulate cortex.

Velocity-selective (VS) magnetization preparation has shown great promise for non-contrast-enhanced (NCE) magnetic resonance angiography (MRA) with the ability to generate positive angiographic contrast directly using a single 3D acquisition. However, existing VS-MRA methods have an issue of aliased saturation around a certain velocity, known as velocity field-of-view (vFOV), which can cause undesired signal loss in arteries. This study aimed to develop a new version of the VS preparation pulse sequence that overcomes the aliased saturation problem in conventional VS preparation. Utilizing the fact that an excitation profile is the Fourier transform of excitation k-space sampling, we sampled the k-space in a non-uniform fashion by scaling gradient pulses accordingly to have aliased excitation diffused over velocity. The variable density sampling function was numerically optimized to maximize the average of the velocity passband signal while minimizing its variance. The optimized variable density VS magnetization was validated through Bloch simulations and applied to peripheral NCE MRA in healthy subjects. The in-vivo experiments showed that the proposed variable density VS-MRA significantly lowered arterial signal loss observed in conventional VS-MRA, as evidenced by a higher arterial signal-to-noise ratio (58.50 ± 14.29 vs. 55.54 ± 12.32; p < 0.05) and improved artery-to-background contrast-to-noise ratio (22.75 ± 7.57 vs. 20.60 ± 6.51; p < 0.05).

Convolutional neural networks (CNNs) have been used widely to predict biological brain age based on brain magnetic resonance (MR) images. However, CNNs focus mainly on spatially local features and their aggregates and barely on the connective information between distant regions. To overcome this issue, we propose a novel multi-hop graph attention (MGA) module that exploits both the local and global connections of image features when combined with CNNs. After insertion between convolutional layers, MGA first converts the convolution-derived feature map into graph-structured data by using patch embedding and embedding-distance-based scoring. Multi-hop connections between the graph nodes are modeled by using the Markov chain process. After performing multi-hop graph attention, MGA re-converts the graph into an updated feature map and transfers it to the next convolutional layer. We combined the MGA module with sSE (spatial squeeze and excitation)-ResNet18 for our final prediction model (MGA-sSE-ResNet18) and performed various hyperparameter evaluations to identify the optimal parameter combinations. With 2788 three-dimensional T1-weighted MR images of healthy subjects, we verified the effectiveness of MGA-sSE-ResNet18 with comparisons to four established, general-purpose CNNs and two representative brain age prediction models. The proposed model yielded an optimal performance with a mean absolute error of 2.822 years and Pearson’s correlation coefficient (PCC) of 0.968, demonstrating the potential of the MGA module to improve the accuracy of brain age prediction.

Velocity-selective (VS) magnetization preparation has shown great promise for non-contrast-enhanced (NCE) magnetic resonance angiography (MRA) with the ability to generate positive angiographic contrast directly using a single 3D acquisition. However, existing VS-MRA methods have an issue of aliased saturation around a certain velocity, known as velocity field-of-view (vFOV), which can cause undesired signal loss in arteries. This study aimed to develop a new version of the VS preparation pulse sequence that overcomes the aliased saturation problem in conventional VS preparation. Utilizing the fact that an excitation profile is the Fourier transform of excitation k-space sampling, we sampled the k-space in a non-uniform fashion by scaling gradient pulses accordingly to have aliased excitation diffused over velocity. The variable density sampling function was numerically optimized to maximize the average of the velocity passband signal while minimizing its variance. The optimized variable density VS magnetization was validated through Bloch simulations and applied to peripheral NCE MRA in healthy subjects. The in-vivo experiments showed that the proposed variable density VS-MRA significantly lowered arterial signal loss observed in conventional VS-MRA, as evidenced by a higher arterial signal-to-noise ratio (58.50 ± 14.29 vs. 55.54 ± 12.32; p < 0.05) and improved artery-to-background contrast-to-noise ratio (22.75 ± 7.57 vs. 20.60 ± 6.51; p < 0.05).

Three-dimensional (3D) projection acquisition (PA) imaging has recently gained attention because of its advantages, such as achievability of very short echo time, less sensitivity to motion, and undersampled acquisition of projections without sacrificing spatial resolution. However, larger subjects require a stronger Nyquist criterion and are more likely to be affected by outer-volume signals outside the field of view (FOV), which significantly degrades the image quality. Here, we proposed a variable slab-selective projection acquisition (VSS-PA) method to mitigate the Nyquist criterion and effectively suppress aliasing streak artifacts in 3D PA imaging. The proposed method involves maintaining the vertical orientation of the slab-selective gradient for frequency-selective spin excitation and the readout gradient for data acquisition. As VSS-PA can selectively excite spins only in the width of the desired FOV in the projection direction during data acquisition, the effective size of the scanned object that determines the Nyquist criterion can be reduced. Additionally, unwanted signals originating from outside the FOV (e.g., aliasing streak artifacts) can be effectively avoided. The mitigation of the Nyquist criterion owing to VSS-PA was theoretically described and confirmed through numerical simulations and phantom and human lung experiments. These experiments further showed that the aliasing streak artifacts were nearly suppressed.

The purpose of the current study was to develop spatially and velocity‐selective (SVS) magnetization preparation pulses for noncontrast‐enhanced peripheral MR angiography (MRA) to provide comparisons with velocity‐selective (VS) MRA with comparison to velocity‐selective (VS). VS preparation pulses were designed by concatenating multiple excitation steps, each of which was a combination of a hard RF pulse, VS unipolar gradient pulses, and refocusing RF pulses. SVS preparation pulses were designed by replacing the hard RF pulse with a sinc‐shaped RF pulse combined with a symmetric tripolar gradient pulse (which does not perturb the velocity encoding by the VS unipolar gradient pulses). Numerical simulations were performed to verify the intended hybrid excitation selectivity of SVS pulses taking account of tissue relaxation, magnetic field errors, and eddy currents. In vivo experiments were performed in healthy subjects to verify the hybrid excitation selectivity, as well as to demonstrate the visualization of the entire peripheral arteries using six‐station protocols. As demonstrated by numerical simulations, SVS preparation yielded a notch‐shaped longitudinal magnetization (M_z)‐velocity response within the spatial stopband (the same as VS preparation) and preserved the M_z of spins outside the stopband, regardless of its velocity. We confirmed these observations also through in vivo tests with good agreement in normalized arterial and muscle signal intensities. In six‐station peripheral MRA experiments, the proposed SVS‐MRA yielded significantly higher arterial signal‐to‐noise ratio (SNR) (51.6 ± 14.3 vs. 38.9 ± 10.9; p < 0.001) and contrast‐to‐noise ratio (CNR) (41.2 ± 13.0 vs. 31.3 ± 10.5; p < 0.001) compared with VS‐MRA. The proposed SVS‐MRA improves arterial SNR and CNR compared with VS‐MRA by mitigating undesired presaturation of arterial blood upstream the imaging field of view.

This study aimed to optimize velocity-selective magnetic resonance angiography (VS-MRA) protocols for whole-neck angiography and demonstrate its feasibility in healthy subjects with comparisons to clinical 3D time-of-flight (TOF) angiography. To help optimize VS-MRA protocols, 2D phase-contrast (PC) flow imaging and 3D B₀ and B₁ field mappings were performed on five healthy volunteers. Based on these measurements, a slab-selective (SS) inversion preparation was applied prior to a VS saturation preparation to further suppress venous blood, while the VS preparation pulse was designed with compensation for field offsets. VS-MRA and 3D TOF were performed on six healthy subjects, and relative contrast ratios (CRs) between artery and muscle signals were calculated for twenty arterial regions for comparisons. The pre-compensated VS pulse improved the visualization of the subclavian arteries and suppression of background tissues, which involved large B₀ and B₁ field errors. The combination of SS and VS preparations effectively suppressed venous blood. While the relative CR values were 0.78 ± 0.08 and 0.72 ± 0.10 for VS-MRA and 3D TOF, respectively, over the twenty segments, VS-MRA outperformed 3D TOF in visualizing arterial segments of a small size or with a horizontal orientation, such as subclavian, facial, and occipital arteries. The proposed neck VS-MRA with the field-error-compensated VS preparation combined with the SS preparation is feasible and superior to 3D TOF in visualizing small and/or horizontally oriented arterial segments.