Kai Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Kai Wang has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- Jing Wang
University of Texas Southwestern Medical Center
11 shared publications Zhiguo Zhou
Department of Biostatistics Data Science, KUMC, United States of America
5 shared publicationsDavid Sher
University of Texas Southwestern Medical Center
3 shared publicationsHui Liu
Department of Biostatistics Data Science, KUMC, United States of America
3 shared publications- Rongfang Wang
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education, School of Artificial Intelligence, Xidian University, Xi’an 710071, China
3 shared publications - You Zhang
University of Texas Southwestern Medical Center
3 shared publications
Latest Publications
- Open Access
- 1 June 2026
- Published by IOP Publishing in Machine Learning: Science and Technology
- Vol. 7 (3) , 035067
Enhancing risk stratification in cancer treatment outcomes: a deep learning-based comparative study of survival and binary models
For cancer treatment outcome prediction, binary classification models categorize patients based on fixed time points, which may oversimplify disease progression. In contrast, survival prediction models leverage time-to-event data, offering a more comprehensive analysis. This study compares the performance of deep learning-based binary and survival prediction models for treatment outcome prediction. We focus on head and neck cancer (HNC) and kidney cancer, two malignancies with distinct treatment paradigms, to assess the advantages of survival modeling over binary classification. Using the RadCure HNC dataset ( n = 2,550), we developed fused convolution neural network and transformer models to predict mortality and recurrence at 2 years post-treatment. For kidney cancer, we used a modified transformer-based multi-instance learning model on the Cancer Genome Atlas Kidney Renal Clear Cell Carcinoma (TCGA-KIRC) dataset ( n = 226), integrating histopathology, multi-omics, and clinical features to predict outcomes at 3 years. Survival models outperformed binary models across single- and multi-modal data. On RadCure dataset, survival models achieved area under the receiver operating characteristic curves (AUROCs) of 0.852, 0.835, 0.811, and 0.828 for mortality and local, regional and distant recurrence predictions, surpassing binary models (0.834, 0.833, 0.809, and 0.805). Notably, survival models for overall survival trained on as little as 20% of the available data achieved performance comparable to binary models trained on the full dataset. With TCGA-KIRC dataset, survival model improved AUROCs to 0.812 (mortality) and 0.836 (recurrence), outperforming binary models (0.770 and 0.798). Overall, survival prediction models provide more robust and data-efficient risk stratification compared to binary models in deep learning setting. By capturing time-dependent risk trajectories, these models enable better data driven personalized cancer treatment strategies.
- 1 June 2026
- Published by IOP Publishing in Machine Learning: Science and Technology
- Vol. 7 (3) , 035067
Open AccessEnhancing risk stratification in cancer treatment outcomes: a deep learning-based comparative study of survival and binary models
For cancer treatment outcome prediction, binary classification models categorize patients based on fixed time points, which may oversimplify disease progression. In contrast, survival prediction models leverage time-to-event data, offering a more comprehensive analysis. This study compares the performance of deep learning-based binary and survival prediction models for treatment outcome prediction. We focus on head and neck cancer (HNC) and kidney cancer, two malignancies with distinct treatment paradigms, to assess the advantages of survival modeling over binary classification. Using the RadCure HNC dataset ( n = 2,550), we developed fused convolution neural network and transformer models to predict mortality and recurrence at 2 years post-treatment. For kidney cancer, we used a modified transformer-based multi-instance learning model on the Cancer Genome Atlas Kidney Renal Clear Cell Carcinoma (TCGA-KIRC) dataset ( n = 226), integrating histopathology, multi-omics, and clinical features to predict outcomes at 3 years. Survival models outperformed binary models across single- and multi-modal data. On RadCure dataset, survival models achieved area under the receiver operating characteristic curves (AUROCs) of 0.852, 0.835, 0.811, and 0.828 for mortality and local, regional and distant recurrence predictions, surpassing binary models (0.834, 0.833, 0.809, and 0.805). Notably, survival models for overall survival trained on as little as 20% of the available data achieved performance comparable to binary models trained on the full dataset. With TCGA-KIRC dataset, survival model improved AUROCs to 0.812 (mortality) and 0.836 (recurrence), outperforming binary models (0.770 and 0.798). Overall, survival prediction models provide more robust and data-efficient risk stratification compared to binary models in deep learning setting. By capturing time-dependent risk trajectories, these models enable better data driven personalized cancer treatment strategies.
- 30 August 2025
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 197 (Pt A) , 110992
Language-guided multimodal domain generalization for outcome prediction of head and neck cancer
- 30 August 2025
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 197 (Pt A) , 110992
Language-guided multimodal domain generalization for outcome prediction of head and neck cancer
- 29 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. 6 (8) , 2027-2040
SAMScore: A Content Structural Similarity Metric for Image Translation Evaluation
Image translation has wide applications, such as style transfer and modality conversion, usually aiming to generate images having both high degrees of realism and faithfulness. These problems remain difficult, especially when it is important to preserve content structures. Traditional image-level similarity metrics are of limited use, since the content structures of an image are high-level, and not strongly governed by pixel-wise faithfulness to an original image. To fill this gap, we introduce SAMScore, a generic content structural similarity metric for evaluating the faithfulness of image translation models. SAMScore is based on the recent high-performance Segment Anything Model (SAM), which allows content similarity comparisons with standout accuracy.We applied SAMScore on 19 image translation tasks, and found that it is able to outperform all other competitive metrics on all tasks. We envision that SAMScore will prove to be a valuable tool that will help to drive the vibrant field of image translation, by allowing for more precise evaluations of new and evolving translation models.
- 29 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. 6 (8) , 2027-2040
SAMScore: A Content Structural Similarity Metric for Image Translation Evaluation
Image translation has wide applications, such as style transfer and modality conversion, usually aiming to generate images having both high degrees of realism and faithfulness. These problems remain difficult, especially when it is important to preserve content structures. Traditional image-level similarity metrics are of limited use, since the content structures of an image are high-level, and not strongly governed by pixel-wise faithfulness to an original image. To fill this gap, we introduce SAMScore, a generic content structural similarity metric for evaluating the faithfulness of image translation models. SAMScore is based on the recent high-performance Segment Anything Model (SAM), which allows content similarity comparisons with standout accuracy.We applied SAMScore on 19 image translation tasks, and found that it is able to outperform all other competitive metrics on all tasks. We envision that SAMScore will prove to be a valuable tool that will help to drive the vibrant field of image translation, by allowing for more precise evaluations of new and evolving translation models.
- 29 November 2024
- Published by Elsevier in Clinical Oncology
- Vol. 38, 103702
Progression-Free Survival Prediction for Locally Advanced Cervical Cancer After Chemoradiotherapy With MRI-based Radiomics
- 29 November 2024
- Published by Elsevier in Clinical Oncology
- Vol. 38, 103702
Progression-Free Survival Prediction for Locally Advanced Cervical Cancer After Chemoradiotherapy With MRI-based Radiomics
- 27 August 2024
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 181, 109050
ASF-LKUNet: Adjacent-scale fusion U-Net with large kernel for multi-organ segmentation
- 27 August 2024
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 181, 109050
ASF-LKUNet: Adjacent-scale fusion U-Net with large kernel for multi-organ segmentation
- 18 March 2023
- Published by Springer Nature
- p. 144-153
Recurrence-Free Survival Prediction Under the Guidance of Automatic Gross Tumor Volume Segmentation for Head and Neck Cancers
- 18 March 2023
- Published by Springer Nature
- p. 144-153
Recurrence-Free Survival Prediction Under the Guidance of Automatic Gross Tumor Volume Segmentation for Head and Neck Cancers
- 10 December 2022
- Published by Wiley in The Laryngoscope
- Vol. 133 (10) , 2754-2760
Joint Vestibular Schwannoma Enlargement Prediction and Segmentation Using a Deep Multi‐task Model
Objective: To develop a deep‐learning‐based multi‐task (DMT) model for joint tumor enlargement prediction (TEP) and automatic tumor segmentation (TS) for vestibular schwannoma (VS) patients using their initial diagnostic contrast‐enhanced T1‐weighted (ceT1) magnetic resonance images (MRIs).Methods: Initial ceT1 MRIs for VS patients meeting the inclusion/exclusion criteria of this study were retrospectively collected. VSs on the initial MRIs and their first follow‐up scans were manually contoured. Tumor volume and enlargement ratio were measured based on expert contours. A DMT model was constructed for jointly TS and TEP. The manually segmented VS volume on the initial scan and the tumor enlargement label (≥20% volumetric growth) were used as the ground truth for training and evaluating the TS and TEP modules, respectively.Results: We performed 5‐fold cross‐validation with the eligible patients (n = 103). Median segmentation dice coefficient, prediction sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were measured and achieved the following values: 84.20%, 0.68, 0.78, 0.72, and 0.77, respectively. The segmentation result is significantly better than the separate TS network (dice coefficient of 83.13%, p = 0.03) and marginally lower than the state‐of‐the‐art segmentation model nnU‐Net (dice coefficient of 86.45%, p = 0.16). The TEP performance is significantly better than the single‐task prediction model (AUC = 0.60, p = 0.01) and marginally better than a radiomics‐based prediction model (AUC = 0.70, p = 0.17).Conclusion: The proposed DMT model is of higher learning efficiency and achieves promising performance on TEP and TS. The proposed technology has the potential to improve VS patient management.Level of Evidence: NA Laryngoscope, 133:2754–2760, 2023
- 10 December 2022
- Published by Wiley in The Laryngoscope
- Vol. 133 (10) , 2754-2760
Joint Vestibular Schwannoma Enlargement Prediction and Segmentation Using a Deep Multi‐task Model
Objective: To develop a deep‐learning‐based multi‐task (DMT) model for joint tumor enlargement prediction (TEP) and automatic tumor segmentation (TS) for vestibular schwannoma (VS) patients using their initial diagnostic contrast‐enhanced T1‐weighted (ceT1) magnetic resonance images (MRIs).Methods: Initial ceT1 MRIs for VS patients meeting the inclusion/exclusion criteria of this study were retrospectively collected. VSs on the initial MRIs and their first follow‐up scans were manually contoured. Tumor volume and enlargement ratio were measured based on expert contours. A DMT model was constructed for jointly TS and TEP. The manually segmented VS volume on the initial scan and the tumor enlargement label (≥20% volumetric growth) were used as the ground truth for training and evaluating the TS and TEP modules, respectively.Results: We performed 5‐fold cross‐validation with the eligible patients (n = 103). Median segmentation dice coefficient, prediction sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were measured and achieved the following values: 84.20%, 0.68, 0.78, 0.72, and 0.77, respectively. The segmentation result is significantly better than the separate TS network (dice coefficient of 83.13%, p = 0.03) and marginally lower than the state‐of‐the‐art segmentation model nnU‐Net (dice coefficient of 86.45%, p = 0.16). The TEP performance is significantly better than the single‐task prediction model (AUC = 0.60, p = 0.01) and marginally better than a radiomics‐based prediction model (AUC = 0.70, p = 0.17).Conclusion: The proposed DMT model is of higher learning efficiency and achieves promising performance on TEP and TS. The proposed technology has the potential to improve VS patient management.Level of Evidence: NA Laryngoscope, 133:2754–2760, 2023
- 29 June 2022
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 67 (13) , 135012
Real-time MRI motion estimation through an unsupervised k-space-driven deformable registration network (KS-RegNet)
Purpose . Real-time three-dimensional (3D) magnetic resonance (MR) imaging is challenging because of slow MR signal acquisition, leading to highly under-sampled k-space data. Here, we proposed a deep learning-based, k-space-driven deformable registration network (KS-RegNet) for real-time 3D MR imaging. By incorporating prior information, KS-RegNet performs a deformable image registration between a fully-sampled prior image and on-board images acquired from highly-under-sampled k-space data, to generate high-quality on-board images for real-time motion tracking. Methods . KS-RegNet is an end-to-end, unsupervised network consisting of an input data generation block, a subsequent U-Net core block, and following operations to compute data fidelity and regularization losses. The input data involved a fully-sampled, complex-valued prior image, and the k-space data of an on-board, real-time MR image (MRI). From the k-space data, under-sampled real-time MRI was reconstructed by the data generation block to input into the U-Net core. In addition, to train the U-Net core to learn the under-sampling artifacts, the k-space data of the prior image was intentionally under-sampled using the same readout trajectory as the real-time MRI, and reconstructed to serve an additional input. The U-Net core predicted a deformation vector field that deforms the prior MRI to on-board real-time MRI. To avoid adverse effects of quantifying image similarity on the artifacts-ridden images, the data fidelity loss of deformation was evaluated directly in k-space. Results . Compared with Elastix and other deep learning network architectures, KS-RegNet demonstrated better and more stable performance. The average (±s.d.) DICE coefficients of KS-RegNet on a cardiac dataset for the 5- , 9- , and 13-spoke k-space acquisitions were 0.884 ± 0.025, 0.889 ± 0.024, and 0.894 ± 0.022, respectively; and the corresponding average (±s.d.) center-of-mass errors (COMEs) were 1.21 ± 1.09, 1.29 ± 1.22, and 1.01 ± 0.86 mm, respectively. KS-RegNet also provided the best performance on an abdominal dataset. Conclusion . KS-RegNet allows real-time MRI generation with sub-second latency. It enables potential real-time MR-guided soft tissue tracking, tumor localization, and radiotherapy plan adaptation.
- 29 June 2022
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 67 (13) , 135012
Real-time MRI motion estimation through an unsupervised k-space-driven deformable registration network (KS-RegNet)
Purpose . Real-time three-dimensional (3D) magnetic resonance (MR) imaging is challenging because of slow MR signal acquisition, leading to highly under-sampled k-space data. Here, we proposed a deep learning-based, k-space-driven deformable registration network (KS-RegNet) for real-time 3D MR imaging. By incorporating prior information, KS-RegNet performs a deformable image registration between a fully-sampled prior image and on-board images acquired from highly-under-sampled k-space data, to generate high-quality on-board images for real-time motion tracking. Methods . KS-RegNet is an end-to-end, unsupervised network consisting of an input data generation block, a subsequent U-Net core block, and following operations to compute data fidelity and regularization losses. The input data involved a fully-sampled, complex-valued prior image, and the k-space data of an on-board, real-time MR image (MRI). From the k-space data, under-sampled real-time MRI was reconstructed by the data generation block to input into the U-Net core. In addition, to train the U-Net core to learn the under-sampling artifacts, the k-space data of the prior image was intentionally under-sampled using the same readout trajectory as the real-time MRI, and reconstructed to serve an additional input. The U-Net core predicted a deformation vector field that deforms the prior MRI to on-board real-time MRI. To avoid adverse effects of quantifying image similarity on the artifacts-ridden images, the data fidelity loss of deformation was evaluated directly in k-space. Results . Compared with Elastix and other deep learning network architectures, KS-RegNet demonstrated better and more stable performance. The average (±s.d.) DICE coefficients of KS-RegNet on a cardiac dataset for the 5- , 9- , and 13-spoke k-space acquisitions were 0.884 ± 0.025, 0.889 ± 0.024, and 0.894 ± 0.022, respectively; and the corresponding average (±s.d.) center-of-mass errors (COMEs) were 1.21 ± 1.09, 1.29 ± 1.22, and 1.01 ± 0.86 mm, respectively. KS-RegNet also provided the best performance on an abdominal dataset. Conclusion . KS-RegNet allows real-time MRI generation with sub-second latency. It enables potential real-time MR-guided soft tissue tracking, tumor localization, and radiotherapy plan adaptation.
- 8 June 2022
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 67 (12) , 125004
Locoregional recurrence prediction in head and neck cancer based on multi-modality and multi-view feature expansion
Objective. Locoregional recurrence (LRR) is one of the leading causes of treatment failure in head and neck (H&N) cancer. Accurately predicting LRR after radiotherapy is essential to achieving better treatment outcomes for patients with H&N cancer through developing personalized treatment strategies. We aim to develop an end-to-end multi-modality and multi-view feature extension method (MMFE) to predict LRR in H&N cancer. Approach. Deep learning (DL) has been widely used for building prediction models and has achieved great success. Nevertheless, 2D-based DL models inherently fail to utilize the contextual information from adjacent slices, while complicated 3D models have a substantially larger number of parameters, which require more training samples, memory and computing resources. In the proposed MMFE scheme, through the multi-view feature expansion and projection dimension reduction operations, we are able to reduce the model complexity while preserving volumetric information. Additionally, we designed a multi-modality convolutional neural network that can be trained in an end-to-end manner and can jointly optimize the use of deep features of CT, PET and clinical data to improve the model’s prediction ability. Main results. The dataset included 206 eligible patients, of which, 49 had LRR while 157 did not. The proposed MMFE method obtained a higher AUC value than the other four methods. The best prediction result was achieved when using all three modalities, which yielded an AUC value of 0.81. Significance. Comparison experiments demonstrated the superior performance of the MMFE as compared to other 2D/3D-DL-based methods. By combining CT, PET and clinical features, the MMFE could potentially identify H&N cancer patients at high risk for LRR such that personalized treatment strategy can be developed accordingly.
- 8 June 2022
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 67 (12) , 125004
Locoregional recurrence prediction in head and neck cancer based on multi-modality and multi-view feature expansion
Objective. Locoregional recurrence (LRR) is one of the leading causes of treatment failure in head and neck (H&N) cancer. Accurately predicting LRR after radiotherapy is essential to achieving better treatment outcomes for patients with H&N cancer through developing personalized treatment strategies. We aim to develop an end-to-end multi-modality and multi-view feature extension method (MMFE) to predict LRR in H&N cancer. Approach. Deep learning (DL) has been widely used for building prediction models and has achieved great success. Nevertheless, 2D-based DL models inherently fail to utilize the contextual information from adjacent slices, while complicated 3D models have a substantially larger number of parameters, which require more training samples, memory and computing resources. In the proposed MMFE scheme, through the multi-view feature expansion and projection dimension reduction operations, we are able to reduce the model complexity while preserving volumetric information. Additionally, we designed a multi-modality convolutional neural network that can be trained in an end-to-end manner and can jointly optimize the use of deep features of CT, PET and clinical data to improve the model’s prediction ability. Main results. The dataset included 206 eligible patients, of which, 49 had LRR while 157 did not. The proposed MMFE method obtained a higher AUC value than the other four methods. The best prediction result was achieved when using all three modalities, which yielded an AUC value of 0.81. Significance. Comparison experiments demonstrated the superior performance of the MMFE as compared to other 2D/3D-DL-based methods. By combining CT, PET and clinical features, the MMFE could potentially identify H&N cancer patients at high risk for LRR such that personalized treatment strategy can be developed accordingly.
- 31 July 2020
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 65 (15) , 155009
Multifaceted radiomics for distant metastasis prediction in head & neck cancer
Accurately predicting distant metastasis in head & neck cancer has the potential to improve patient survival by allowing early treatment intensification with systemic therapy for high-risk patients. By extracting large amounts of quantitative features and mining them, radiomics has achieved success in predicting treatment outcomes for various diseases. However, there are several challenges associated with conventional radiomic approaches, including: (1) how to optimally combine information extracted from multiple modalities; (2) how to construct models emphasizing different objectives for different clinical applications; and (3) how to utilize and fuse output obtained by multiple classifiers. To overcome these challenges, we propose a unified model termed as multifaceted radiomics (M-radiomics). In M-radiomics, a deep learning with stacked sparse autoencoder is first utilized to fuse features extracted from different modalities into one representation feature set. A multi-objective optimization model is then introduced into M-radiomics where probability-based objective functions are designed to maximize the similarity between the probability output and the true label vector. Finally, M-radiomics employs multiple base classifiers to get a diverse Pareto-optimal model set and then fuses the output probabilities of all the Pareto-optimal models through an evidential reasoning rule fusion (ERRF) strategy in the testing stage to obtain the final output probability. Experimental results show that M-radiomics with the stacked autoencoder outperforms the model without the autoencoder. M-radiomics obtained more accurate results with a better balance between sensitivity and specificity than other single-objective or single-classifier-based models.
- 31 July 2020
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 65 (15) , 155009
Multifaceted radiomics for distant metastasis prediction in head & neck cancer
Accurately predicting distant metastasis in head & neck cancer has the potential to improve patient survival by allowing early treatment intensification with systemic therapy for high-risk patients. By extracting large amounts of quantitative features and mining them, radiomics has achieved success in predicting treatment outcomes for various diseases. However, there are several challenges associated with conventional radiomic approaches, including: (1) how to optimally combine information extracted from multiple modalities; (2) how to construct models emphasizing different objectives for different clinical applications; and (3) how to utilize and fuse output obtained by multiple classifiers. To overcome these challenges, we propose a unified model termed as multifaceted radiomics (M-radiomics). In M-radiomics, a deep learning with stacked sparse autoencoder is first utilized to fuse features extracted from different modalities into one representation feature set. A multi-objective optimization model is then introduced into M-radiomics where probability-based objective functions are designed to maximize the similarity between the probability output and the true label vector. Finally, M-radiomics employs multiple base classifiers to get a diverse Pareto-optimal model set and then fuses the output probabilities of all the Pareto-optimal models through an evidential reasoning rule fusion (ERRF) strategy in the testing stage to obtain the final output probability. Experimental results show that M-radiomics with the stacked autoencoder outperforms the model without the autoencoder. M-radiomics obtained more accurate results with a better balance between sensitivity and specificity than other single-objective or single-classifier-based models.