LZLei Zhang

Lei Zhang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Lei Zhang has a total of 90 co-authors. Here are the top co-authors by number of shared publications:

  • Huimao Zhang

    Department of Radiology, The First Hospital of Jilin University, Changchun, China

    4 shared publications
  • HongChao Wang

    Department of Radiology, Jilin Provincial Key Laboratory of Medical Imaging and Big Data, Radiology and Technology Innovation Center of Jilin Province, Jilin Provincial International Joint Research Center of Medical Artificial Intelligence, The First Hospital of Jilin University, Changchun, People’s Republic of China

    3 shared publications
  • Huimao Zhang

    Department of Radiology, First Affiliated Hospital of Jilin University, Changchun, China

    3 shared publications
  • Alfred Wei Chieh Kow

    Division of Hepatobiliary & Pancreatic Surgery, Department of Surgery, National University of Singapore, Singapore

    2 shared publications
  • BingYang Bian

    Department of Radiology, Jilin Provincial Key Laboratory of Medical Imaging and Big Data, Radiology and Technology Innovation Center of Jilin Province, Jilin Provincial International Joint Research Center of Medical Artificial Intelligence, The First Hospital of Jilin University, Changchun, People’s Republic of China

    2 shared publications
  • Guoyue Lv

    Department of Hepatobiliary and Pancreatic Surgery, General Surgery Center, First Hospital of Jilin University, Changchun, China

    2 shared publications

Latest Publications

  • Open Access

    ProtoSurv: prototype-guided adaptation of computed tomography foundation model for lung-cancer prognosis prediction

    by Chengcai Liu, Shitian Li, Haolin Sang, Xingyu Huang, Chenghao Wang, Yifei Shu, Yi Wu, Jie Tian, Huimao Zhang, Lei Zhang, et al.

    Prognosis prediction is important for precision treatment in lung cancer. Chest computed tomography (CT) and deep learning are feasible approaches for learning imaging-based prognostic representations. However, prognostic labels are difficult to obtain because survival endpoints require long-term follow-up; thus, labeled data for supervised training is limited, increasing the risk of overfitting. Vision-language pretraining offers an approach to mitigate this problem by leveraging large-scale unlabeled CT data; however, most existing methods are not specifically designed to learn prognosis-related features. We propose ProtoSurv, a prototype-guided learning framework for adapting large-scale pretrained CT representations for lung-cancer prognosis prediction under limited supervision. ProtoSurv first pretrains a CT-report vision-language model using 409,261 chest CT sequences from 104,783 patients. It then uses limited labeled prognosis data to construct class prototypes in the pretrained feature space. These prototypes guide pseudolabel assignment, confidence-based filtering, and feature calibration, allowing the selection and refinement of task-relevant unlabeled samples before downstream prognosis modeling. We evaluated ProtoSurv based on two lung-cancer prognosis tasks: progression-free survival prediction in 507 patients receiving targeted therapy and overall survival prediction in 420 patients receiving (chemo-)radiotherapy. ProtoSurv achieved area under the curve/concordance index values of 0.765/0.684 and 0.822/0.727 for the first and second datasets, respectively, outperforming the conventional clinical models and representative deep learning baselines. These results suggest that the prototype-guided adaptation can improve the use of large-scale unlabeled CT data for prognosis modeling when labeled survival data are limited.

  • Open Access

    Preoperative 3D morphological phenotyping of hepatocellular carcinoma via a foundation model: a non-invasive biomarker for recurrence risk stratification

    by Hongbin Wang, Xiaogen Zhou, Kun Zhang, Kan He, Wenao Ma, Mingyang Li, Han Xue, Ye Sun, Yingzhu Cui, Xiaopeng Wang, et al.

  • Open Access

    Integrated model based on ultrasound attenuation and metabolic biomarkers for noninvasive assessment of hepatic fat fraction categories in MASLD: a QCT-referenced study

    by Zejun Ma, Hongbin Wang, Xiaojie Sun, Shuyu Zhou, Ye Sun, Xiaoguang Wang, Tao Li, Xiaohe Yang, Jiancheng Xu, Wanjun Guo, et al.

    Background: Hepatic steatosis is a common metabolic disorder for which accessible noninvasive assessment remains clinically relevant. This study aimed to evaluate the QCT-referenced performance of ultrasound attenuation (USAT) and an integrated model combining metabolic biomarkers for non-invasive categorization according to hepatic fat fraction in metabolic dysfunction-associated steatotic liver disease (MASLD), using quantitative computed tomography (QCT)-derived hepatic fat fraction categories as pragmatic imaging comparator labels. Methods: A total of 172 participants were enrolled and categorized into hepatic fat fraction categories by QCT. USAT values, along with serum levels of ALT, AST, and ferritin, were collected. Three models were evaluated: USAT-only, laboratory-only, and an integrated USAT + laboratory model. Model performance was assessed using fold-separated internal validation, including nested five-fold stratified cross-validation with feature selection performed within training folds. Harrell’s optimism-corrected bootstrap analysis was also performed as a supplementary internal validation method. Results: USAT values increased significantly with hepatic fat fraction categories (P<0.001). In fold-separated internal validation, the USAT-only model achieved an AUC of 0.847 (95% CI, 0.780–0.902) for QCT-referenced detection of imaging-defined steatosis, while the laboratory-only model achieved an AUC of 0.753 (95% CI, 0.665–0.829). A fixed integrated model including USAT, ALT, and ferritin achieved an AUC of 0.845 (95% CI, 0.772–0.906), but did not significantly improve AUC compared with USAT alone. Multiclass categorization remained exploratory and limited, particularly for Category 2, which included only 31 participants and showed weak discrimination in the corrected random forest analysis (AUC=0.569 for the USAT + ALT + ferritin model). Subgroup analyses showed higher performance in females, younger participants, and those with higher BMI. Conclusion: USAT shows promise as a noninvasive adjunct for QCT-referenced detection of imaging-defined hepatic steatosis in a single-center Chinese health-examination cohort, pending external validation. Adding ALT and ferritin may improve sensitivity and calibration, but did not significantly improve AUC over USAT alone. Because discrimination for the intermediate Category 2 group remained weak, the current model should not be considered reliable for full hepatic fat fraction categorization. Further validation in external, multicenter, and more diverse populations is required before any clinical use.

  • Fine-Tuned Large Language Model for Automated Radiology Impression Generation: A Multicenter Evaluation

    by Mingyang Li, Yaning Wang, Zheng Miao, Jiaqi Gong, Simin Yang, Han Xue, Qi Yang, Lijun Duan, Lin Mu, Ying Mu, et al.

    Large-scale fine-tuning produced a highly accurate large language model for automated radiology impression generation, demonstrating a validated approach to improve reporting efficiency and consistency across multicenter examinations.


  • Open Access

    High-resolution structural magnetic resonance examination of the Habenula in patients with first-episode depression: an exploratory radiomics diagnostic value analysis based on cluster analysis

    by Lin Hou, BingYang Bian, ShuXin Luan, XingChen Pan, MingYang Li, Han Xue, HongChao Wang, HuiMao Zhang, Lei Zhang

    The habenula (Hb) is a vital hub for the monoaminergic pathway and plays a crucial role in depression pathophysiology. However, owing to its small size and heterogeneity between individuals, there is no consensus on imaging alterations in the Hb in depression. This study aimed to examine the differences in the Hb between healthy controls (HCs) and patients with first-episode depression (FED) who were not taking any antidepressants, and to assess the value of Hb voxel cluster radiomic features in discriminating patients with FED from HCs. This cross-sectional study included 94 participants (47 HCs and 47 patients with FED) who underwent 3-T magnetic resonance imaging. Differences in the Hb volume and T1 values between the two groups were examined. Correlations among volume, T1 value, depression severity, and age were also examined. Furthermore, a clustering-based radiomics model to differentiate patients with FED from HCs was developed and validated. In HCs, the Hb T1 value was positively related to age, whereas that of patients with FED showed no significant correlation. The prediction performance was improved in the clustering-based radiomics model (area under the curve [AUC] = 0.844) compared with the traditional model (AUC = 0.708). Our findings imply that the Hb and its internal heterogeneity are imaging markers for depression studies. Not applicable.


  • Open Access

    Visualizing the Habenula Using 3T High-Resolution MP2RAGE and QSM: A Preliminary Study

    by BingYang Bian, Lin Hou, YaTing Chai, YueLuan Jiang, XingChen Pan, Yang Sun, HongChao Wang, DongDong Qiu, ZeChen Yu, Hua Zhao, et al.

    BACKGROUND AND PURPOSE: The habenula is a key node in the regulation of emotion-related behavior. Accurate visualization of the habenula and its reliable quantitative analysis is vital for the assessment of psychiatric disorders. To obtain high-contrast habenula images and allow them to be compatible with clinical applications, this preliminary study compared 3T MP2RAGE and quantitative susceptibility mapping with MPRAGE by evaluating the habenula segmentation performance. MATERIALS AND METHODS: Ten healthy volunteers were scanned twice with 3T MPRAGE and MP2RAGE and once with quantitative susceptibility mapping. Image quality and visibility of habenula anatomic features were analyzed by 3 radiologists using a 5-point scale. Contrast assessments of the habenula and thalamus were also performed. The reproducibility of the habenula volume from MPRAGE and MP2RAGE was evaluated by manual segmentation and the Multiple Automatically Generated Template brain segmentation algorithm (MAGeTbrain). T1 values and susceptibility were measured in the whole habenula and habenula geometric subregion using MP2RAGE T1-mapping and quantitative susceptibility mapping. RESULTS: The 3T MP2RAGE and quantitative susceptibility mapping demonstrated clear boundaries and anatomic features of the habenula compared with MPRAGE, with a higher SNR and contrast-to-noise ratio (all P < .05). Additionally, 3T MP2RAGE provided reliable habenula manual and MAGeTbrain segmentation volume estimates with greater reproducibility. T1-mapping derived from MP2RAGE was highly reliable, and susceptibility contrast was highly nonuniform within the habenula. CONCLUSIONS: We identified an optimized sequence combination (3T MP2RAGE combined with quantitative susceptibility mapping) that may be useful for enhancing habenula visualization and yielding more reliable quantitative data.


  • Open Access

    Recent Advances in Habenula Imaging Technology: A Comprehensive Review

    by BingYang Bian, Bei Zhang, ChinTing Wong, Le Dou, XingChen Pan, HongChao Wang, ShiYu Guo, HuiMao Zhang, Lei Zhang

    The habenula (Hb) is involved in many natural human behaviors, and the relevance of its alterations in size and neural activity to several psychiatric disorders and addictive behaviors has been presumed and investigated in recent years using magnetic resonance imaging (MRI). Although the Hb is small, an increasing number of studies have overcome the difficulties in MRI. Conventional structural‐based imaging also has great defects in observing the Hb contrast with adjacent structures. In addition, more and more attention should be paid to the Hb's functional, structural, and quantitative imaging studies. Several advanced MRI methods have recently been employed in clinical studies to explore the Hb and its involvement in psychiatric diseases. This review summarizes the anatomy and function of the human Hb; moreover, it focuses on exploring the human Hb with noninvasive MRI approaches, highlighting strategies to overcome the poor contrast with adjacent structures and the need for multiparametric MRI to develop imaging markers for diagnosis and treatment follow‐up.Level of Evidence: 3Technical Efficacy Stage: 2


  • Open Access

    Computed Tomography‐Based Radiomic Analysis for Preoperatively Predicting the Macrovesicular Steatosis Grade in Cadaveric Donor Liver Transplantation

    by Shengnan Ding, Weimin Yang, Xiaodong Sun, Yan Guo, Guangjie Zhao, Jinzhu Yang, Lei Zhang, Guoyue Lv

    This study is aimed at determining the ability of computed tomography‐ (CT‐) based radiomic analysis to distinguish between grade 0/1 and grade 2/3 macrovesicular steatosis (MaS) in cadaveric donor liver transplantation cases. Preoperative noncontrast‐enhanced CT images of 150 patients with biopsy‐confirmed MaS were analyzed retrospectively; these patients were classified into the low‐grade MaS (n = 100, grade 0 or 1) and high‐grade MaS (n = 50, grade 2 or 3) groups. Three‐dimensional spherical regions of interest of 40 pixel (2.5 cm) in diameter were placed in the right anterior and left lateral segments of the liver. Thereafter, 300 regions of interest (ROIs) were segmented and randomly assigned to the training and testing groups at a ratio of 7 : 3. A total of 402 radiomic features were extracted from each ROI. For MaS classification, a radiomic model was established using multivariate logistic regression analysis. Clinical data, including age, sex, and liver function, were collected to establish the clinical model at the patient level. The performances of the radiomic and clinical models, i.e., the diagnostic discrimination, calibration, and clinical utilities, were evaluated. The radiomic model, with seven selected features, depicted a good discrimination with an area under the receiver operating characteristic curve (AUC) of 0.907 (95% confidence interval (CI): 0.869–0.940) in the training cohort and 0.906 (95% CI: 0.843–0.959) in the testing cohort. The calibration curve revealed good agreement between the predicted and observed probabilities in the training and testing cohorts (both P > 0.05 in the H‐L test). Decision curve analysis revealed that the radiomic model was more beneficial than the treat‐all or treat‐none schemes for predicting the MaS grade. Alanine transaminase and gamma‐glutamyl transferase were used for building the clinical model, and the AUC was 0.784 in the total cohort. The CT‐based radiomic model outperforming the conventional clinical model could provide an important reference for MaS grading in cadaveric liver donors.