To construct an automatic segmentation model to segment female urine control anatomy on MRI images based on deep learning methods in order to improve the segmentation efficiency and accuracy.

A dataset comprising 49 female pelvic floor muscle MRI images [30 women with varying degrees of pelvic organ prolapse (POP) and 19 healthy individuals], obtained from Faculty of Biomedical Engineering and Medical Imaging in Army Medical University, was used for model training and testing. The dataset was split into a training set (17 normal cases and 22 POP cases) and a testing set (4 normal cases and 6 POP cases) in a ratio of 8∶2. The training set was used to train UNet, UNet+++, Dense UNet, and UNet++ models separately, and then input into each network. The model achieving the highest testing accuracy was selected as the backbone network.

Under the training of UNet, UNet+++, Dense UNet, and UNet++, the 4 models achieved average Dice similarity coefficients of 61.82%, 57.94%, 57.63%, and 62.76%, respectively, for the segmentation of 5 anatomical structures (compressor urethrae, urethra sphincter body, bladder wall, bladder cavity and urethra submucosa). The corresponding Intersection over Union (IoU) score was 49.74%, 46.59%, 46.07%, and 49.44%, while the accuracy rate was 61.74%, 55.03%, 59.23%, and 61.91%, respectively for the 4 models. Notably, UNet++ consistently outperformed UNet, UNet+++, and Dense UNet across the 3 metrics, indicating that UNet++ achieved the highest overall segmentation accuracy.

In UNet, UNet++, Dense UNet and UNet++ for automatic segmentation of 5 female urine control anatomical elements, UNet++ achieves the best overall segmentation accuracy.

To establish and validate a multimodal MRI radiomics model based on intratumoral and peritumoral heterogeneity for prediction of spatial pattern of locally recurrent high-grade gliomas (HGGs).

A retrospective analysis was conducted on the clinical and imaging data of all HGGs patients who underwent maximum safe resection followed by postoperative radiotherapy combined with temozolomide treatment and experienced in local recurrence in Army Medical Center of PLA from 2012 to 2021. Two radiologists independently assessed the spatial patterns of locally recurrence HGGs through continuous follow-up MRI data, and primarily categorized the pattern into intra-resection cavity recurrence and extra-resection cavity recurrence. The subjected patients were randomly divided into a training set and a validation set in a 7∶3 ratio. In the training set, Pearson or Spearman correlation analysis and least absolute shrinkage and selection operator (LASSO) analysis were employed to screen radiomic features within the intratumoral and peritumoral regions, as well as to calculate radiomic scores. A radiomics model was established using logistic regression analysis. The performance of the model was assessed using calibration curves, Hosmer-Lemeshow goodness-of-fit test, and the area under the receiver operating characteristic curve (AUC). Validation of the model was performed in the validation set.

A total of 121 patients with locally recurrent HGGs were enrolled in this study, including 54 in intra-resection cavity recurrence group and 67 in extra-resection cavity recurrence group. Among them, 84 were assigned into the training set and 37 into the validation set. In the training set, the radiomics score for the extra-resection cavity recurrence group was 0.424 (0.278, 0.573), which was higher than that for the intra-resection cavity recurrence group [-0.030 (-0.226,0.248), P<0.001]. In the validation set, the radiomics score for the extra-resection cavity recurrence group was 0.369 (0.258, 0.487), which was higher than that for the intra-resection cavity recurrence group [0.277 (0.103, 0.322), P=0.033]. The established radiomics model exhibited good calibration and performed well in predicting spatial recurrence patterns, with an AUC value of 0.844 (95%CI: 0.749~0.914) in the training set and 0.706 (95%CI:0.534~0.844) in the validation set.

Our multimodal radiomics model combined with intratumoral and peritumoral heterogeneity can predict the spatial pattern of locally recurrent HGGs, providing a basis for individualized treatment of HGGs.

To develop a CT radiomics-based prediction model for prognosis of pancreatic ductal adenocarcinoma (PDAC) in order to provide evidence for individualized treatment decisions.

A retrospective study was carried on 118 PDAC patients admitted in the First Affiliated Hospital of Army Medical University between January 2020 and December 2023. They were assigned into a training group (n=83) and a validation group (n=35) at a 7∶3 ratio. ITK-SNAP software was used to perform 3-D segmentation on the preoperatively enhanced arterial phase CT images, and radiomic features were extracted using pyradiomics. High-reproducibility features were selected through ICC analysis (>0.85), and core features were determined using LASSO regression to construct the Rad-score. Cox regression analysis was employed to develop both a radiomics model and a model integrating radiomic and clinical features for predicting overall survival in PDAC patients. Receiver operating characteristic (ROC) curves and calibration curves were plotted to evaluate the prognostic models for survival prediction.

From 1 453 extracted radiomic features, 7 core features were finally selected to construct the Rad-score. The radiomics prediction model based on the Rad-score achieved an AUC value of 0.796 (95%CI: 0.702~0.890) and 0.744 (95%CI: 0.589~0.899) for 1-year survival prediction in the training and validation groups, respectively. The integrated model combining 2 types of features together demonstrated improved performance with an AUC value of 0.906 (95%CI: 0.842~0.970) and 0.872 (95%CI: 0.753~0.992) in the 2 groups. Calibration curve analysis indicated good prediction accuracy for both models.

Both the CT radiomics-based model and the integrated model incorporating clinical features demonstrate good predictive performance for survival outcomes.

To develop an automatic segmentation network for thyroid nodules by integrating wavelet transform and CNN-Transformer in order to improve the efficiency and accuracy of ultrasound image segmentation.

A total of 1 371 sets of ultrasound images of thyroid nodules were collected from Department of Ultrasonography of Second Affiliated Hospital of Army Medical University between May 2023 and February 2024. After preprocessing and normalization, the data were divided into training, validation, and testing sets in a ratio of 8∶1∶1. Based on UNet, CNN and Swin-Transformer were used in parallel as the encoder, with a wavelet transform module inserted between the encoder and decoder to construct a thyroid nodule segmentation network. The performance of the segmentation model was evaluated on the collected internal dataset using accuracy, IoU, and Dice coefficient metrics.

The finally verified 1 371 sets of ultrasonic thyroid nodules had an average Dice coefficient of 79.63% and an IoU of 67.30%. Compared with UNet, the segmentation accuracy was improved by 1.02%. The segmentation model obtained accurate location and smooth edges of thyroid nodules, and the segmentation was more consistent in thyroid nodule edge and morphology with those marked by doctors manually when compared with other segmentations. Compared with UNet, this segmentation method can learn the texture of nodules more fully and avoid the situation that nodules had been incorrectly divided into surrounding tissues.

Our developed segmentation model based on wavelet transform and CNN-Transformer demonstrates better segmentation accuracy in comparison to conventional UNet variants, such as UNet, Attention-UNet, and UNetv2, and medical segment anything models like SAM Med2D. This segmentation method enables accurate segmentation of ultrasound thyroid nodules, thereby enhancing clinical workflow efficiency through automated precise delineation.

To develop a machine learning model integrating preoperative chest CT radiomic features with clinical data for predicting 5-year postoperative recurrence risk in stage Ⅰ non-small cell lung cancer (NSCLC) patients undergoing surgical resection.

A total of 217 patients with pathologically confirmed stage Ⅰ NSCLC (selected from 778 initially screened cases based on our inclusion and exclusion criteria) treated in Army Medical Center of PLA between January 2014 and December 2019 were retrospectively enrolled, including 53 recurrence cases and 164 non-recurrence cases within 5-year follow-up. They were randomly divided into a training set (n=173) and a validation set (n=44) in a ratio of 8:2. Radiomic models were established based on extracted features from tumor-dominant regions of interest (ROI) on CT images, while clinical models were developed using demographic characteristics and preoperative laboratory examinations. A combined model was further constructed by integrating both feature sets, and model performance was compared to identify the optimal predictive model.

This study screened the features from non-contrast CT images and ultimately selected 7 radiomic features for constructing radiomic model. Among 6 machine learning algorithms, the adaptive boosting (Adaboost) model demonstrated the best overall predictive performance, with an area under the curve (AUC) of 0.866 (95% CI: 0.808~0.923; accuracy: 0.832, specificity: 0.884) in the training set and of 0.806 (95% CI: 0.630~0.983; accuracy: 0.795, specificity: 0.971) in the validation set. Univariate and multivariate logistic regression analyses identified 4 clinical features for clinical model construction. The clinical model achieved an AUC value of 0.874 (95% CI: 0.821~0.928; accuracy: 0.827, specificity: 0.891) in the training set and 0.813 (95% CI: 0.677~0.948; accuracy: 0.636, specificity: 0.600) in the validation set. By integrating the 7 radiomic features and 4 clinical features using a feature-level fusion strategy, the combined model exhibited further improved predictive performance, with an AUC value of 0.953 (95% CI: 0.924~0.983; accuracy: 0.884, specificity: 0.860) and 0.852 (95% CI: 0.729~0.976; accuracy: 0.682, specificity: 0.629), respectively in the training set and the validation set.

The combined model integrating preoperative CT radiomic features with clinical risk factors may provide an evidence-based framework for evaluating 5-year postoperative recurrence risk in stage Ⅰ NSCLC patients.

To investigate the mechanism by which hypoxia-inducible factor (HIF)-1α and circ-UBE2G1 mutually regulate and promote thyroid cancer (THCA) metastasis under hypoxic microenvironment.

‍The GEPIA database was used to analyze the expression characteristics and correlation between circ-UBE2G1 and HIF-1α in THCA. The relationship of circ-UBE2G1 and miR-330-3p with survival rate was analyzed using Kaplan-Meier survival curve. After THCA cells were exposed to hypoxia, HIF-1α was silenced by transfection to analyze its regulation for circ-UBE2G1 transcription. The targeting relationship between miR-330-3p and either circ-UBE2G1 or HIF-1α was verified by sequence prediction and dual luciferase reporter assay. The effects of HIF-1α overexpression and circ-UBE2G1 silencing on THCA cell migration and invasion were analyzed after corresponding transfections. A tumor-bearing nude mouse model was established by subcutaneous injection of THCA cells with HIF-1α overexpression and circ-UBE2G1 silencing, respectively. THCA tissues and adjacent normal samples were clinically collected to analyze the expression levels and correlations of circ-UBE2G1, miR-330-3p, and HIF-1α.

Bioinformatics analysis showed that circ-UBE2G1 was positively correlated with HIF-1α (P<0.01), and its high expression was associated with a low survival rate in THCA patients (P=0.024). Inhibition of HIF-1α blocked the promotive effect of hypoxia on circ-UBE2G1 (P<0.05). Silencing circ-UBE2G1 inhibited the migration and invasion of THCA cells (P<0.05), and reversed the promotive effect of HIF-1α on these processes (P<0.05). Dual luciferase reporter gene assay revealed that miR-330-3p targeted both circ-UBE2G1 and HIF-1α (P<0.05). Silencing circ-UBE2G1 led to increased levels of miR-330-3p (P<0.05), whereas increasing miR-330-3p inhibited the expression of HIF-1α (P<0.05). In clinical THCA samples, both circ-UBE2G1 and HIF-1α were increased and positively correlated (P<0.05), while miR-330-3p was lowly expressed and negatively correlated with both circ-UBE2G1 and HIF-1α (P<0.05).

Hypoxia promotes the transcription of circ-UBE2G1 by inducing HIF-1α expression, and circ-UBE2G1 promotes HIF-1α expression by targeting miR-330-3p, thereby promoting THCA metastasis.

To investigate the synergistic therapeutic effects of HER2 antibody-drug conjugate (HER2-ADC) combined with anti-PD-1 antibody (anti-PD-1) on HER2-expressing bladder cancer and elucidate its regulatory mechanisms on the tumor immune microenvironment.

Orthotopic tumor models were established in 40 female C57BL/6 mice (6~8 weeks old, body mass 18~22 g) using MB49 bladder cancer cells overexpressing human HER2. When tumors reached 100 mm³, the mice were randomized into (n=10) control (intraperitoneal injection of 1.0 mL PBS), anti-PD-1 (200 µg per mouse every 3 d), HER2-ADC (2.5 mg/kg once weekly), and combination groups (same regimens as above monotherapy). Tumor volume and body mass were measured every 3 d during 28-day treatment. Tumor growth kinetics and survival rates were analyzed. Post-treatment survival was monitored until natural death to determine median survival time (n=5). At day 28, blood and tumor samples (n=5) were collected to detect myeloid-derived suppressor cells (MDSCs; CD11b⁺Gr1⁺) and regulatory T cells (Tregs; CD4⁺CD25⁺FOXP3⁺) with flow cytometry, tumor-infiltrating CD3⁺T, CD8⁺T, and FOXP3⁺T cells with immunohistochemical assasy, and liver/kidney functions [alanine aminotransferase (ALT), aspartate aminotransferase (AST), blood urea nitrogen (BUN), creatinine (CRE)] and tissue damage indicators [lactate dehydrogenase isoenzyme (LDH-L)].

In 28 d after treatment, the combination group obtained significantly smallest tumor volume than the control group and the 2 monotherapy groups (all P<0.01). The longest median survival was observed in the combination group (65 d, P<0.01), followed by the HER2-ADC group (55 d), anti-PD-1 group (53 d) and control group (41 d). After 28 d of treatment, the combination group exhibited obviously the smallest peripheral proportions of MDSCs/Tregs, most tumor-infiltrating CD3⁺T/CD8⁺T cells, and less FOXP3⁺T cells when compared with the 2 monotherapy groups and control group (all P<0.05). While, the 2 monotherapy groups had smaller MDSCs/Tregs proportions than the control group (P<0.05). No significant differences were observed among the 4 groups in serum ALT, AST, BUN, CRE, or LDH-L levels, and all of them were within normal ranges.

HER2-ADC combined with anti-PD-1 suppresses the growth of orthotopic bladder tumor, probably through their synergic effects on down-regulating MDSCs/Treg and enhancing CD8⁺T cell infiltration.

To investigate the effect of β-arrestin1 (ARRB1) on cell proliferation, migration and invasion in oral squamous cell carcinoma (OSCC).

Based on The Cancer Genome Atlas (TCGA) database, the expression profiles of ARRB1 in OSCC were analyzed, and then Gene Set Enrichment Analysis (GSEA) was used to suggest the possible signaling pathways involved, and to explore its potential impact on the prognosis of OSCC patients. Immuinohistochemistry (IHC) was performed to detect the expression of ARRB1 in OSCC tumor tissues and adjacent tissues, and the correlation between ARRB1 expression and clinicopathological features was statistically analyzed. The expression profiles of ARRB1 in SCC-15, CAL-27 and HOK cell lines were verified by qPCR and Western blotting. The ARRB1 overexpression plasmid model was constructed, and its effects on the proliferation, migration and invasion of OSCC cells were analyzed by clone formation, EdU, scratch and Transwell assays.

TCGA showed that the expression level of ARRB1 was significantly lower in head and neck squamous cell carcinoma (HNSC) and OSCC tissues than the corresponding normal tissues (P<0.01). The expression of ARRB1 in OSCC tissues was correlated with tumor differentiation, lymph node metastasis and TNM stage (P <0.05). The OSCC patients with high expression of ARRB1 had a lower survival rate than those with low expression (P<0.01), which was consistent with the results of bioinformatics analysis. The expression level of ARRB1 in SCC-15 and CAL-27 cells was lower than that of HOK cells (P<0.01), and its overexpression significantly inhibited cell proliferation (P<0.05), migration (P<0.01) and invasion (P<0.01).

ARRB1 is lowly expressed in OSCC, its overexpression inhibits the proliferation, migration and invasion of OSCC cells, and it is related to prognosis improvement.

To explore the occurrence of acute kidney injury (AKI) in patients with Stanford type A aortic dissection (TAAD) after surgical treatment, analyze the risk factors, and construct a prediction model.

A retrospective case-control study was conducted on 138 TAAD patients undergoing surgical treatment in our hospital from January 2016 to June 2024. After Kidney Disease Improving Global Outcomes (KDIGO) criteria was performed within 1 week after surgery, they were divided into AKI (n=95) and non-AKI (n=43) groups. Univariate and multivariate logistic regression analyses were performed to identify the perioperative risk factors. Then a nomogram model were constructed, and receiver operating characteristic (ROC) curve was plotted to analyze its predictive efficacy.

The incidence of postoperative AKI was 68.84% (95 cases) in the TAAD patients, including 51.58% (49 cases) of stage 3 AKI and 38.95% (37 cases) requiring continuous renal replacement therapy. The length of ICU stay, time to extubation, and abandonment of treatment were significantly higher in the AKI group than the non-AKI group (P<0.001). Multifactorial analysis showed that the monocyte count on postoperative day 1 (OR=3.521) and preoperative creatinine level (OR=1.019) were independent risk factors for AKI, postoperative uric acid level (OR=1.005) was correlated with AKI, and intraoperative urine volume (OR=0.739) and globulin level at 1 d postoperatively (OR=0.781) were protective factors. The area under the ROC curve of the constructed model was 0.866, with a sensitivity of 0.811 and a specificity of 0.791.

Postoperative AKI occurrence can be reduced in TAAD patients by optimizing the intraoperative urine output, modulating the postoperative inflammatory response, and strengthening nutritional support. Our prediction model for AKI risk is of significance for early clinical identification of high-risk patients in TAAD patients after surgical treatment.

To analyze the temporal trend of thyroid cancer incidence in China from 1992 to 2021 and evaluate the impact of age, period and cohort on the incidence.

Based on the Global Burden of Disease 2021, the trend changes of age-standardized incidence rate (ASIR) of thyroid cancer in China from the period were analyzed, and the obtained data were further analyzed with Joinpoint regression and age-period-cohort analysis. Auto regressive integrated moving average (ARIMA) model was applied to predict the incidence of thyroid cancer in China in the next 20 years.

The ASIR of thyroid cancer among males was increased from 0.70/100 000 to 2.11/100 000 between 1992 and 2021, with an average annual percent change (AAPC) of 3.93% (95%CI: 3.86%~4.00%, P<0.001); For the females, the ASIR was elevated from 1.91/100 000 to 2.87/100 000 in the time, with an AAPC of 1.43% (95%CI: 1.36%~1.50%, P<0.001). On the age effect, the overall variation in ASIR among the females was minor, and that in the males reached its peak in the ≥85-years-old group, with a rate of 230.51/100 000. Regarding the period effect, the risk of thyroid cancer incidence in the males and females changed with the period and exhibited a temporal increase, and the incidence risk was consistently higher in the males than the females across different periods. Taking the incidence risk during 1992 and 1996 as a reference (RR=1), the males (RR=2.94, 95%CI: 2.78~3.11) and the females (RR=1.37, 95%CI: 1.32~1.42) had the highest risk of thyroid cancer in the period from 2017 to 2021. On the cohort effect, the risk of thyroid cancer in both sexes showed an increasing trend with cohort change. When taking the cohort during 1992 and 1996 as a reference (RR=1), the females in the pre-1986 cohort exhibited a higher risk of incidence than the males, but in the post-1996 cohort, the males demonstrated a greater risk of incidence than the females. The males in the 2012-2016 cohort had the highest risk of thyroid cancer when compared to the reference group (RR=2.54, 95%CI: 1.80~3.57), and the females in the cohort from 1992 to 2016 had higher risk of thyroid cancer than the reference group, though no statistical differences were observed. The ARIMA model prediction analysis found that the ASIR of thyroid cancer in males and females in China would continue to increase from 2022 to 2041, and the ASIR of thyroid cancer in males would reach 2.79/100 000 (95% CI: 1.96~3.61), and the ASIR of thyroid cancer in females would reach 3.54/100 000 (95% CI: 3.02~4.06) by 2041.

Between 1992 and 2021, the ASIR of thyroid cancer in China has exhibited a consistent upward trend, with the females having a higher ASIR than the males. However, the risk of developing thyroid cancer is higher in men who are older and those who were born more recently. In the next 20 years, the incidence of thyroid cancer among both men and women in China is projected to continue to rise, and the situation for prevention and control will be challenging.

To investigate the current status of pelvic floor muscle training (PFMT) adherence for preventing low anterior resection syndrome (LARS) in rectal cancer patients with prophylactic ostomy and analyze its influencing factors.

A total of 247 patients who underwent preventive stoma reversal surgery after sphincter-preserving rectal cancer resection between January 2024 and February 2025 at 22 tertiary hospitals in the Beijing-Tianjin-Hebei region were recruited in this study. Their clinical data were collected through a general information questionnaire, a LARS Knowledge-Attitude-Practice and Needs questionnaire, a PFMT Self-Efficacy Scale, and a PFMT Adherence Questionnaire. Univariate analysis, correlation analysis, LASSO regression, and random forest importance ranking were applied to screen the variables, and multiple linear regression analysis was conducted to analyze the significant variables.

The overall PFMT adherence score was 14.52±4.18 among the 247 patients. The random forest algorithm identified 7 key predictors when the minimum error was achieved at a λ value of 2.293. The top 7 variables in importance ranking were PFMT self-efficacy, total knowledge-attitude-practice (KAP) score, education level, primary caregiver, tumor location, stoma reversal time, and chemoradiotherapy. Multiple linear regression analysis revealed that PFMT self-efficacy (P<0.001, β=0.007, 95%CI: 0.004~0.009), total KAP score (P<0.001, β=0.052, 95%CI: 0.035~0.070), stoma reversal time (P<0.030, β=-0.539, 95%CI: -1.025~-0.053), and chemoradiotherapy (P<0.045, β=-0.451, 95%CI: 0.010~0.892) were significant related factors of PFMT adherence (P<0.05).

PFMT adherence for LARS prevention is at a moderate level in rectal cancer patients with prophylactic ostomy. Key factors such as PFMT self-efficacy, total KAP score, stoma reversal time, and chemoradiotherapy are significantly correlated with PFMT adherence.

To develop a prediction model for placenta previa in early pregnancy by analyzing embryo implantation sites and related clinical factors, in order to provide an objective basis for early risk identification.

A retrospective cohort study was conducted on 232 singleton pregnant women delivering in our hospital between September 2020 and March 2024. According to the final pre-delivery ultrasound findings, they were divided into placenta previa group (n=78) and a non-placenta previa group (n=154). Their ultrasound parameters [distance from lower gestational sac margin to cervical os, implantation site (lower/middle-upper uterine segment), and implantation position (anterior/posterior wall)], and clinical data [age, gravidity (categorized as <2 or ≥2 pregnancies), parity (<2 or ≥2 deliveries), and history of intrauterine procedures were collected through electronic medical records. Univariate analysis was used to screen potential predictors (P<0.1), and multivariate logistic regression analysis was employed to identify the predictors (P<0.05) for placenta previa. Then a nomogram prediction model was constructed, which was internally validated with Bootstrap (1 000 bootstrap resamples) and assessed for discrimination with area under the receiver operating characteristic curve (AUC) and for calibration with Hosmer-Lemeshow goodness-of-fit test.

The placenta previa group showed significantly advanced age (≥35 years), lower education level (≤high school), multigravidity (≥2 pregnancies), multiparity (≥2 deliveries), more intrauterine procedures, distance of gestational sac-to-cervical os <5.5 mm, and larger proportion of lower uterine segment implantation than the non-placenta previa group (all P<0.1). But there were no statistical differences between the 2 groups in proportion of posterior wall implantation, history of cesarean section or assisted reproductive technology (ART). Multivariate logistic regression analysis confirmed lower uterine segment implantation (OR=40.40, 95%CI: 14.68~136.19, P<0.001), posterior wall implantation (OR=2.73, 95%CI:1.27~6.28, P=0.013), and intrauterine procedures (OR=3.48, 95%CI: 1.65~7.70, P=0.001) as independent risk factors. The model based on these predictors demonstrated excellent discrimination (AUC value=0.84, 95%CI: 0.79~0.90) and calibration (Hosmer-Lemeshow test Chi-square=3.455, P=0.750).

Lower uterine segment/posterior wall implantation and intrauterine procedures are independent risk factors for placenta previa. Our nomogram model based on these factors shows good predictive efficiency, and can provide reference for early recognition of pregnant woman with high-risk placenta previa.

To develop and validate a deep learning model based on chest CT images to accurately distinguish between drug-resistant (DR-TB) and -sensitive tuberculosis (DS-TB).

A retrospective study was conducted on 722 cases of confirmed secondary tuberculosis admitted in our center from January 2019 to December 2022. According to the results of antimicrobial susceptibility test, they were divided into 357 DS-TB cases and 365 DR-TB cases. Pre-existing U-Net segmentation model was employed to segment the lung parenchyma regions in CT images. The dataset was randomly partitioned into a training set and a testing set in an 8:2 ratio. Six 3D deep learning architectures (3D Swin Transformer, 3D ShuffleNet v2, 3D ViT, 3D MobileNet v2, 3D DenseNet, and 3D ResNet18) were employed to evaluate the discriminative efficiency between DS-TB and DR-TB. Hyperparameters were optimized by five-fold cross-validation on the training set to construct the optimal model. The performance of the constructed model was assessed using area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score. Six radiologists independently evaluated DR-TB identification on the test set, and their performance was compared with the best-performing deep learning model.

The AUC value in DR-TB prediction was 0.583, 0.704, 0.698, 0.758, 0.736, and 0.841, respectively, for 3D Swin Transformer, 3D ShuffleNet v2, 3D ViT, 3D MobileNet v2, 3D DenseNet, and 3D ResNet18. The 3D ResNet18 model demonstrated optimal performance, achieving a sensitivity of 0.935 (95%CI: 0.880~0.987), a specificity of 0.642 (95%CI: 0.492~0.757), a PPV of 0.750 (95%CI: 0.663~0.835), an NPV of 0.896 (95%CI: 0.809~0.976), an AUC value of 0.841, and a F1-score of 0.832. The radiologists got a F1-score of 0.571, 0.450, 0.675, 0.623, 0.617 and 0.635, respectively, and the F1-score of the 3D ResNet18 model is all higher than that of the radiologists. The highest-performing radiologist achieved sensitivity, specificity, PPV and NPV of 0.701 (95%CI: 0.605~0.802), 0.567 (95%CI: 0.447~0.684), 0.651 (95%CI: 0.549~0.757), and 0.623 (95%CI: 0.500~0.754), with all these values lower than those of the 3D ResNet18 model (P<0.05). Class activation mapping showed that the 3D ResNet18 model could focus on key lesion areas. The class activation mapping demonstrated that the 3D ResNet18 model could effectively focus on critical lesion regions.

Our 3D ResNet18 model shows the best predictive performance in identifying DR-TB, and is expected to assist clinical diagnosis for DR-TB.