Publications

The proposed domain-adaptive framework for myocardial segmentation uses cine MR data from various manufacturers and employs a class-imbalance self-training structure to address domain shift issues between different scanner systems. The framework iteratively refines the segmentation model using pseudo labels generated in the target domain with a class-specific label threshold to re-balance pixel inconsistencies between different label regions, demonstrating improved segmentation quality compared to other methods when tested on adapting a UNet trained with Siemens scanner data to effectively segment Philips scanner data.

F. Xing, X. Liu, I. Aganj, P. Kim, B. W. Choi, J. Woo

SPIE Medical Imaging (2025)

The study proposes a semi-supervised local anomaly detection method using mask inpainting models to effectively identify bone marrow lesions (BMLs) in high-resolution knee MRIs, demonstrating superior performance compared to existing methods. By integrating 3D femur bone segmentation, mask inpainting, and post-processing techniques, the research shows promising results in BML detection, with performance significantly improving on higher-resolution images and potentially advancing imaging-based osteoarthritis research.

S. Qin, M. Zhang, J. Shan, T. Shin, J. Woo, F. Xing

SPIE Medical Imaging (2025)

Contrast-enhanced T1-weighted MRI using gadolinium-based contrast agents has been crucial for nasopharyngeal carcinoma diagnosis, but safety concerns have prompted research into alternative methods. This study developed a diagnostic model using non-contrast images that achieved high accuracy comparable to contrast-enhanced imaging, suggesting a promising contrast-free approach for NPC detection.

Z. Li et al.

SPIE Medical Imaging (2025)-Robert F. Wagner All-Conference Best Student Paper Award (Finalist)

Multi-source Black-box Domain Adaptation using Label space-induced Pseudo label Refinement (LPR) addresses privacy and intellectual property concerns by utilizing only API predictions from multiple source domains, without requiring access to source data or model parameters. The proposed framework introduces a Pseudo label Refinery Network that learns inter-domain relationships through warm-up and refinement phases, with experimental results demonstrating competitive performance compared to state-of-the-art approaches across four benchmark datasets.

C. Yoo, X. Liu, F. Xing, J. Woo and J.-W. Kang

IEEE Trans on Image Processing (2025)

This study proposes an unsupervised framework for coronary lesion detection in computed tomography coronary angiography (CTCA) that uses vessel linearization to generate small linear volume patches and trains a deep autoencoder model exclusively on normal patches to detect anomalies without requiring lesion location labels. Results demonstrate high accuracy on a public dataset with 97.5% accuracy using per-patient average reconstruction error and 85% using the number of patches with high reconstruction error, offering a scalable solution for coronary artery disease detection without relying on large labeled datasets.

D. Kang, J. Choi, J. Woo, C.-C. J. Kuo

IEEE ISBI (2025)

The proposed data-efficient multi-task framework leverages hierarchical segmentation features within advanced Swin UNETR to predict glioblastoma patient survival, integrating multi-scale features to capture detailed spatial information while maintaining computational efficiency. Through segmentation pre-training, feature fine-tuning aligned with survival prediction, and incorporation of age information, the model achieves superior segmentation accuracy and state-of-the-art survival prediction performance on the BraTS dataset, offering a robust solution for clinical prognosis.

K. Kim, F. Xing, H.-J. Kong, E. Santarnecchi, H. Shih, T. Bortfeld, G. El Fakhri, X. Liu, J.-H. Choi, J. Woo

IEEE ISBI (2025)

This study developed a deep learning-based diffusion model to automatically and accurately delineate gross tumor volumes (GTVs) for soft tissue sarcomas using multi-modality medical images. By training on expert-labeled data and modeling inter-reader variability through confidence maps, the model achieved competitive performance across key metrics compared to U-Net models.

Y. Dong et al.

Physics and Imaging in Radiation Oncology (2025)

This work proposes an improved dynamic speech MRI method using Poisson-disc sampling combined with a locally higher-rank model and regional-optimized temporal basis, which better represents the distinct temporal characteristics of oral and static regions. This approach significantly improves spatiotemporal image quality while reducing overall acquisition time by 50%, achieving 3D dynamic speech MRI with 2-mm isotropic spatial resolution.

R. Jin, Y. Li, R. K. Shosted, F. Xing, I. Gilbert, J. L. Perry, J. Woo, Z.-P. Liang, B. P. Sutton

Magnetic Resonance in Medicine (2024)

This study introduces an iterative framework to improve point-supervised medical image segmentation by converting point inputs into refined bounding box suggestions and leveraging the MedSAM model’s generalizability. By using a semantic box-prompt generator and a prompt-guided spatial refinement module, we demonstrated superior performance in whole brain tumor segmentation compared to traditional point-supervised segmentation methods.

X. Liu et al.

IEEE MIC (2024)

We propose a treatment-conditioned regression model to predict glioblastoma patient survival time using preoperative MR scans and treatment information. Unlike previous approaches that ignored treatment effects, our model incorporates treatment types (GTR, STR, or no resection) using adaptive instance normalization within each convolutional layer. This unified approach improves survival prediction accuracy and enables personalized treatment planning, as demonstrated on the BraTS20 dataset.

X. Liu, N. Shusharina, H. A. Shih, C.-C. J. Kuo, G. El Fakhri, J. Woo

SPIE Medical Imaging (2024)

The human tongue moves through a coordinated sequence of internal muscles working together, with five specific muscles being key to pushing the tongue forward (protrusion). Using specialized MRI techniques and statistical analysis, our team discovered that these muscles work in a predictable sequence to create tongue protrusion, advancing our understanding of how the tongue functions.

H. Park et al.

JASA EL (2024)

The paper evaluates state-of-the-art few-shot object detection methods for detecting anatomical structures in medical imaging, specifically head and neck anatomy in CT scans. It highlights the challenges posed by variable appearances and limited annotated datasets in medical imaging, motivating the use of few-shot approaches like QA-FewDet, Meta Faster R-CNN, and FCT. Experimental results demonstrate that these methods can accurately detect anatomical structures using novel datasets with few images (1- to 30-shot), showing promising potential for clinical integration.

K. Lee et al.

SPIE Medical Imaging (2024) - Student Travel Award

This study explores variations in velar consonant production across different phonological environments using dynamic magnetic resonance imaging (MRI).

F. Xing, J. Zhuo, M. Stone, X. Liu, T. G. Reese, V. J. Wedeen, J. L. Prince, J. Woo

JASA (2024)

The proposed framework for tagged-to-cine MR sequence synthesis leverages both spatial and temporal information through a light spatial-temporal transformer (LiST2) that exploits local and global attention with relatively lightweight parameters while incorporating directional product relative position-time bias for spatial-temporal correlation awareness. Additionally, a recurrent sliding fine-tuning scheme enhances temporal consistency, resulting in superior performance over comparison methods when evaluated on paired tagged and cine MRI sequences.

X. Liu, F. Xing, Z. Bian, T. Arias-Vergara, P. Andrea Pérez-Toro, A. Maier, M. Stone, J. Zhuo, J. L. Prince and J. Woo

MICCAI (2024)

An off-the-shelf model for unsupervised domain adaptation (OSUDA) can be vulnerable to backdoor attacks due to the inability to access or control source domain data, but a proposed method quantifies backdoor sensitivity using a Lipschitz constant and eliminates the infection by overwriting backdoor-related channel kernels, validated on a multi-center cardiac dataset.

X. Liu, F. Xing, H. Gaggin, C.-C. J. Kuo, G. El Fakhri, J. Woo

IEEE ISBI (2024)

In this work, we develop an unsupervised domain adaptation approach to carry out class-wise conditional alignment.

X. Liu, F. Xing, J. You, J. Lu, C.-C. J. Kuo, G. El Fakhri, J. Woo

IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2024)

Magnetic Resonance Imaging with tagging (tMRI) has long been utilized for quantifying tissue motion and strain during deformation. However, a phenomenon known as tag fading, a gradual decrease in tag visibility over time, often complicates post-processing.

Z. Bian, A. Alshareef, S. Wei, J. Chen, Y. Wang, J. Woo, D. L. Pham, J. Zhuo, A. Carass, J. L. Prince

SPIE Medical Imaging (2024)–Best Student Paper Award

We propose a contrastive learning approach that improves the detection of phonological classes from MRI data when acoustic signals are unavailable, raising frame-wise recognition from an f1 of 0.74 to 0.85. Applied to clinical assessments in tongue cancer patients, this method demonstrates promising results for speech disorder evaluation, potentially enhancing patient outcomes.

T. Arias-Vergara, P. Andrea Perez-Toro, X. Liu, F. Xing, M. Stone, J. Zhuo, J. L. Prince, M. Schuster, E. Noth, J. Woo, A. Maier

Interspeech (2024)

Magnetic Resonance Imaging with tagging (tMRI) has long been utilized for quantifying tissue motion and strain during deformation. However, a phenomenon known as tag fading, a gradual decrease in tag visibility over time, often complicates post-processing.

J. Cho, X. Liu, F. Xing, J. Ouyang, G. El Fakhri, J. Park, J. Woo

SPIE Medical Imaging (2024)

With the rapid development of natural language processing (NLP) technologies in recent years, chatbots using large language models (LLMs) have received much attention nowadays. A comprehensive overview on bias and fairness in chatbot systems is given in this paper.

J. Xue, Y.-C. Wang, C. Wei, X. Liu, J. Woo, C.-C. J. Kuo

APSIPA Trans on Signal and Information Processing (2024)

This study investigates the potential role of 18F-FDG PET in reducing intra- and inter-reader variability thereby improving reproducibility of GTV delineation in STS, without incurring additional costs or radiation exposure.

E. Najem, T. Marin, Y. Zhuo, R. Maria Lahoud, F. Tian, A. Beddok, L. Rozenblum, F. Xing, M. Moteabbed, R. Lim, X. Liu, J. Woo, S. John Lostetter, A. Lamane, Y.-L. Chen, C. Ma, G. El Fakhri

Radiotherapy and Oncology (2024)

In this work, we develop a new deep learning-based image denoising especially to deal with rapid fluorescence and ultrasound captures where traditional noise mitigation strategies are limited, such as increasing pixel dwell time or frame averaging.

K. Lee, H. Lee, M. Lee, J. Chang, C.-C. Jay Kuo, S.-J. Oh, J. Woo, J. Y. Hwang

APSIPA Trans on Signal and Information Processing (2024)

In this work, we develop a noise-robust sleep stage network via adversarial training with an auxiliary model.

C. H. Yoo, X. Liu, F. Xing, G. El Fakhri, J. Woo, J.-W. Kang

IEEE Trans on Biomedical Engineering (2023)

This paper presents a novel unsupervised phase-based 3D motion estimation technique for tagged MRI that addresses traditional challenges through two key innovations, including a sinusoidal transformation of harmonic phase input and a Jacobian determinant-based learning objective for incompressibility.

J. Bian et al.

MIDL (2023)

In this work, we seek to progressively evolve an “off-the-shelf” trained segmentation model to diverse datasets with additional anatomical categories in a unified manner.

X. Liu, H. A. Shih, F. Xing, E. Santarnecchi, G. El Fakhri, J. Woo

MICCAI (2023)

The Stochastic Confidence Network (SCN) is a novel deep learning model designed to handle outliers in medical imaging, ensuring accurate and robust diagnoses. Experimental results on databases for breast tumors and Alzheimer’s disease demonstrated that SCN significantly improves diagnostic performance by eliminating irrelevant outlier data and resampling outliers into a typical distribution, outperforming state-of-the-art models.

K. Lee, H. Lee, G. El Fakhri, J. Sepulcre, X. Liu, F. Xing, J. Y. Hwang, J. Woo

MICCAI MTSAIL (2023)

We introduce a highly innovative imaging method to study the complex velopharyngeal system and introduce the potential future clinical applications of a VP atlas in cleft care.

J. Perry, I. Gilbert, F. Xing, R. Jin, D. Kuehn, R. Shosted, J. Woo, Z. P. Liang, B. Sutton

The Cleft Palate Craniofacial Journal (2023)

We developed successive subspace learning for cardiac disease classification from cine MRI.

X. Liu, F. Xing, H. K. Gaggin, C.-C. J. Kuo, G. El Fakhri, J. Woo

IEEE ISBI (2023)

We developed a new self-supervised domain adaptive segmentation method for breast cancer.

K. Lee, H. Lee, G. El Fakhri, J. Woo, J. Y. Hwang

MICCAI (2023)

We propose off-the-shelf (OS) UDA (OSUDA), aimed at image segmentation, by adapting an OS segmentor trained in a source domain to a target domain, in the absence of source domain data in adaptation.

X. Liu, F. Xing, G. El Fakhri, J. Woo

Medical Image Analysis (2023)

In this work, we developed a generative self-training framework for domain adaptive image translation with continuous value prediction and regression objectives.

X. Liu, J. L. Prince, F. Xing, J. Zhuo, T. Reese, M. Stone, G. El Fakhria, J. Woo

Medical Image Analysis (2023)

The proposed plastic light transformer (PLT) framework translates tongue motion weighting maps derived from tagged MRI to corresponding audio waveforms through directional product relative position bias and single-level spatial pyramid pooling, enabling flexible processing of variable-sized inputs without information loss. To enhance spectrogram realism despite limited training samples, the framework employs pair-wise utterance consistency with Maximum Mean Discrepancy constraint and adversarial training, outperforming conventional models when tested on 29 subjects speaking two utterances.

X. Liu, F. Xing, M. Stone, J. Zhuo, S. Fels, J. L. Prince, G. El Fakhri, J. Woo

MICCAI (2023)

In this work, we developed a federated learning approach for personalized skin diagnosis via fine-tuning network.

K. Lee, H. Lee, T. Coutinho Cavalcanti, S. Kim, G. El Fakhri, D. H. Lee, J. Woo, J. Y. Hwang

MICCAI (2023)

We use inverse atlas tongue modeling for investigating motor control similarity.

U. Maity, F. Xing, J. Prince, M. Stone, G. El Fakhri, J. Woo, S. Fels

Interspeech (2023)

In this work, we developed a method for deep learning based inpainting and applied to construct brain atlases.

F. Xing, X. Liu, C.-C. J. Kuo, G. El Fakhri, J. Woo

IEEE J. Biomed. Health Informatics (2022)

This study focuses on achieving near-isotropic 3D dynamic imaging with high frame rates through sparse temporal sampling and region-of-interest based coil compression, resulting in 1.875 × 1.875 × 2 mm³ spatial resolution with 35.6 fps temporal resolution across a 64-mm coverage.

R. Jin, R. K. Shosted, F. Xing, I. Gilbert, J. L. Perry, J. Woo, Z.-P. Liang, B. P. Sutton

Magnetic Resonance in Medicine (2022)

This study addresses the lack of systematic guidelines for determining optimal sample sizes in data augmentation by proposing an automated pipeline using generative adversarial networks (GAN) to enhance diagnostic performance with limited datasets. By demonstrating improved diagnostic parameters in chronic sinusitis radiographs, the research offers a novel approach to overcome training data limitations and potentially assist radiologists in more accurate diagnoses.

H.-J. Kong et al.

Scientific Reports (2022)

This study explores using diffusion tensor MRI to derive muscle fiber orientations for defining clinical target volume boundaries in soft-tissue sarcoma, leveraging machine learning to auto-segment thigh muscles and develop an anisotropic modeling approach. By comparing fiber orientations across different muscles and subjects, the researchers demonstrated a promising method for more precise tumor boundary delineation that could improve treatment planning for sarcoma patients.

Shusharina et al.

Physics in Medicine and Biology (2022)

Unsupervised domain adaptation (UDA) transfers knowledge from a labeled source domain to unlabeled target domains, addressing the challenge of labeling in new domains. To overcome privacy concerns in cross-center collaborations, we propose a UDA framework using a black-box segmentation model with knowledge distillation and unsupervised entropy minimization, validated on multiple datasets and demonstrating potential in clinical settings.

X. Liu, C. Yoo, F. Xing, C.-C. J. Kuo, G. El Fakhri, J.-W. Kang and J. Woo

Frontiers in Neuroscience (2022)

In this work, we developed a lightweight and interpretable machine learning framework for ALS disease classification from structural MRI.

X. Liu, F. Xing, C. Yang, C.-C. J. Kuo, S. Babu, G. El Fakhri, T. Jenkins, J. Woo

IEEE J. Biomed. Health Informatics (2022)

This review summairzes state-of-the-art deep unsupervised domain adaptation techniques that are proposed in recent years.

X. Liu, C. Yoo, F. Xing, H. Oh, G. El Fakhri, J.-W. Kang, J. Woo

APSIPA Trans on Signal and Information Processing (2022)

This paper proposes a novel variational inference framework that models the distribution of plausible segmentation maps in MR images to account for inter-observer variability among radiologists without compromising accuracy. Experimental results using QUBIQ brain growth MRI datasets with seven annotators demonstrate how this approach effectively captures the natural ambiguity in medical image segmentation by encoding multi-reader variability in a latent vector and optimizing a variational autoencoder network.

X. Liu, F. Xing, T. Marine, G. El Fakhri, J. Woo

SPIE Medical Imaging (2022)

This research proposes a subtype-aware dynamic unsupervised domain adaptation approach that improves cross-domain alignment by discovering and leveraging fine-grained subtype structures within classes, enforcing both class-wise separation and subtype-wise compactness. The method adaptively identifies subtypes with or without prior information on subtype numbers, recognizing that unlabeled subtypes within a class can diverge from each other while maintaining local proximity, achieving promising results on medical diagnosis tasks.

X. Liu et al.

AAAI (2021)

This work proposed an adversarial unsupervised domain adaptation with conditional and label shift.

X. Liu, Z. Guo, S. Li, F. Xing, J. You, C.-C. J. Kuo, G. El Fakhri, J. Woo

ICCV (2021)

This work proposed severity-aware semantic segmentation with reinforced Wasserstein training.

X. Liu, W. Ji, J. You, G. El Fakhri, J. Woo

CVPR (2020)