Tien Hung Bui has a total of 37 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    END-TO-END MULTI-ISOTOPE ACTIVITY QUANTIFICATION IN ENVIRONMENTAL SOIL SAMPLES USING 1D RESIDUAL NEURAL NETWORKS AND HPGE GAMMA-RAY SPECTROMETRY

    by Ha Ngoc Bui, Duong Thuy Tran, Dinh Thuy Mai, Tran Ngoc Toan, Tat Thang Nguyen, Bui Tien Hung, Truong Minh Anh

    Environmental radioactivity monitoring plays an important role in quantifying radiation exposure risks from natural radionuclides. Traditional spectral analysis procedures still have limitations regarding processing time and subjectivity due to human error. This study develops a customized analysis pipeline that applies a one-dimensional residual neural network (1D-ResNet) to directly quantify multi-isotope activity concentrations from 500 soil samples in Vietnam. By integrating the Savitzky-Golay smoothing technique and physics-informed data augmentation methods (MixUp and Poisson simulation), the model has solved the problem of training data scarcity. Prediction results on the test set show that 1D-ResNet outperforms shallow learning models (MLR, ANN) when handling complex decay chains, achieving a coefficient of determination () of 0.873 for and 0.859 for . The application of the Monte Carlo dropout technique has provided an uncertainty evaluation framework that meets the GUM guidelines for radiation metrology, achieving an empirical coverage probability (ECP) consistent with the nominal 95% confidence interval. However, explainability analysis using attention weight maps has revealed an issue: purely data-driven deep learning networks tend to shortcut learning by using Compton scattering regions for inference rather than photopeaks. From this discovery, the study lays the foundation for developing physics-informed neural networks (PINNs) to ensure future calibration transparency and traceability.

  • Open Access

    Benchmarking Deep Learning for NSCLC PET/CT Segmentation on a Histologically Confirmed Vietnamese Dataset: Validation and Generalization

    by Quang Tuan Ho, Ngoc Ha Bui, Thuy Duong Tran, Quang Huy Khuat, Ngoc Toan Tran, Xuan Chung Le, Huu Quyet Nguyen, Tat Thang Nguyen, Van Thai Nguyen, Dinh Thuy Mai, et al.

    Accurate segmentation of non-small cell lung cancer (NSCLC) on positron emission tomography/computed tomography (PET/CT) is an essential prerequisite for automated metabolic tumor volume (MTV) quantification and staging. Although deep learning models achieve high performance on large-scale datasets, their generalization across different clinical domains is limited by variations in imaging protocols and patient demographics. This study aims to evaluate several deep learning architectures and investigate a transfer learning strategy to mitigate domain shift. Three architectures, including ResNet-backbone 3D U-Net, nnU-Net v2, and Swin UNETR, were benchmarked from scratch and compared with a fine-tuned nnU-Net initialized with AutoPET II weights. Results on the internal dataset showed that the fine-tuned nnU-Net achieved a Dice similarity coefficient (DSC) of 83.4 ± 6.5%, a 95% Hausdorff distance (HD95) of 5.1 ± 3.6 mm, and a precision of 89.6 ± 8.2%. Compared to the nnU-Net v2, the fine-tuned nnU-Net improved the absolute DSC by 6.8% while reducing local training time by 37.5% by bypassing the initial feature-learning phase. The fine-tuned nnU-Net model also demonstrated a high correlation between the MTV and the ground truth (Pearson r = 0.96, p < 0.001), indicating its potential as a reliable automated approach for quantitative MTV extraction and NSCLC prognostic-related analysis.

  • Open Access

    Benchmarking Deep Learning for NSCLC PET/CT Segmentation on a Histologically Confirmed Vietnamese Dataset: Validation and Generalization

    by Quang Tuan Ho, Ngoc Ha Bui, Duong Tran Thuy, Quang Huy Khuat, Ngoc Toan Tran, Xuan Chung Le, Huu Quyet Nguyen, Tat Thang Nguyen, Van Thai Nguyen, Dinh Thuy Mai, et al.

    Accurate segmentation of non-small cell lung cancer (NSCLC) on positron emission tomography/computed tomography (PET/CT) is a critical foundation for automated metabolic tumor volume (MTV) quantification and staging. Although deep learning models achieve high performance on large-scale datasets, their generalization across different clinical domains remains challenged by variations in imaging protocols and patient demographics. This study aims to evaluate several deep learning architectures and investigate a transfer learning strategy to mitigate domain shift on a dataset of 400 PET/CT scans from Vietnamese NSCLC patient cohort. Three architectures, including 3D U-Net, nnU-Net v2, and Swin UNETR, were benchmarked from scratch and compared with a fine-tuned nnU-Net initialized with AutoPET II weights. Results on the internal dataset showed that the fine-tuned nnU-Net achieved a Dice similarity coefficient (DSC) of 83.4 ± 6.5%, a 95% Hausdorff distance (HD95) of 5.1 ± 3.6 mm, and precision of 89.6 ± 8.2%. Compared to the nnU-Net v2, the fine-tuned nnU-Net improved the absolute DSC by 6.8% while reducing the training time by 37.5%. The fine-tuned nnU-Net model also demonstrated a good correlation between the MTV and the ground truth (Pearson r = 0.96, p < 0.001), indicating its potential as a reliable automated approach for quantitative MTV extraction and NSCLC prognostic related analysis.

  • Natural radioactivity and GIS mapping of soils in Binh Phuoc, Vietnam

    by Thuy Duong Tran, Ngoc Ha Bui, Dinh Thuy Mai, Tien Hung Bui, Tat Thang Nguyen, Hai Quan Ho

  • Open Access

    Simulation of Leksell Gamma Knife-4C System with Different Phantoms Using PHITS and Geant4

    by B. T. Hung, T. T. Duong, B. N. Ha

    This study used PHITS and Geant4 code packages to simulate a Leksell Gamma Knife system in order to determine radiation dose distribution in two types of phantoms. The results observed in the water phantom with configurations of single source and 201 sources are in good accord with the prior research, including both simulation and experiment. Several characteristics of Leksell Gamma Knife 4C, such as dose profiles, output factor, FWHM, and penumbra size, are calculated based on Monte Carlo simulations, which show the best consistency with other results. The output factors for collimators of 14 mm, 8 mm, and 4 mm are 0.984, 0.949, and 0.872, respectively. The simulation results with an adult mesh-type reference phantom reveal considerable similarities with the established radiosurgery plans. It indicates that the absorbed dose in brain tumors was highest when utilizing the 18 mm collimator and subsequently reduced with collimator size to 0.65, 0.25, and 0.5 with the 14 mm, 8 mm, and 4 mm collimators, respectively. The absorbed dose has a very low value for other essential organs and decreases with distance from the brain tumor. These findings may explain why the dose to organs decreases linearly as target distance, volume, and collimator size increase.


  • RESEARCH ON THE CONSTRUCTION OF A PHANTOM IN DOSE CALCULATION BY PHITS SIMULATION PROGRAM

    by Hung Bui Tien, Duong Tran Thuy, Ha Bui Ngoc, Thao Nguyen Thi

    In the field of medical physics, determining the dosage for organs is a very complex problem because it is impossible to directly measure the dosage inside a living person's body. Therefore, in order to estimate the dosage, one usually has to use phantoms. For dosage determination and safety assessment, Monte Carlo simulation programs and today's phantoms have become indispensable tools. There are already several standard phantoms in the world but they do not represent biological characteristics by region or for a specific object. Another method used to build characteristic phantoms is done through computed tomography (CT) images and the Hounsfield index. The method requires to convert the grayscale value of each pixel element into the Hounsfield index. As such, it can only be applied to CT imaging. To overcome this shortcoming, this study has developed a program based on a new process that can convert images to geometric format for MCNP\PHITS. The new procedure that we propose will allow researchers to describe the anatomical structures more accurately and be specific to each patient.

  • Open Access

    Detection of Spoilage in Canned Pasteurized Milk Using the Radiographic Imaging Technique

    by D. T. Thuy, H. B. Tien, H. B. Ngoc, T. T. Kim, T. T. Ngoc

    Detection of Spoilage in Canned Pasteurized Milk Using the Radiographic Imaging Technique

  • Evaluation of the effect of filters on reconstructed image quality from cone beam CT system

    by Ngoc Ha Bui, Tien Hung Bui, Thuy Duong Tran, Kim Tuan Tran, Ngoc Toan Tran

    : 3D Filtered Back Projection (FBP) is a three-dimensional reconstruction algorithm usually used in Cone Beam Computed Tomography (CBCT) system. FBP is one of the most popular algorithms due to its reconstruction is fast while quality of the result is acceptable. It can also handle a more considerable amount of data with same computer performance with other algorithms. However, the quality of a reconstructed image by the FBP algorithm strongly depends on spatial filters and denoising filters applied to projections. In this paper an evaluation of the reconstructed image quality of the CBCT system by using different denoising filters and spatial filters to find out the best filters for the CBCT system is performed. The result shows that, there is a significantly decrease of the noise of projection with the combination of Median and Gaussian filters. The reconstructed image has high resolution with Cosine filter and becomes more sharpen with Hanning filter.