RPS 116 - Innovations in AI applications for oncology

Author Block: L. Machado, A. Prat, H. Philippe, T. Danielou, K. Le Floch, M. Ronot, D. Tordjman, P. Manceron, P. Hérent; Paris/FR
Purpose: RECIST 1.1 remains the reference for longitudinal assessment of solid tumors but is slow, labor-intensive, and prone to inter-reader variability. Most AI tools are unimodal, organ-specific, and restricted to single timepoints, limiting clinical use. We present a pan-tumor, CT/MRI lesion-tracking pipeline for semi-automated longitudinal follow-up using foundation models.
Methods or Background: To achieve lesion tracking, we used features from Curia, a multimodal foundation model (Dancette et al., 2025, arXiv). From the baseline lesion, Curia enables the generation of a bounding-box on the follow-up exam, localizing the follow-up lesion. This bounding-box is then processed by the segmentation model Oncopilot-v2 to obtain the follow-up mask. Oncopilot-v2 is a fine-tuning of Raps3D (Danielou et al., 2025, arXiv) on 16,035 lesions (e.g., liver, lung, brain, lymph nodes) in CT and MRI, following the same procedure as Oncopilot (Machado et al., 2025, NPJ Precision Oncology).
Evaluation used a multicenter dataset of 238 patients with 391 baseline target lesions (mean 1.64/patient; CT/MRI ratio 1.15; mean follow-up 290 days). Longest diameter and short axis were computed on baseline and follow-up masks, and RECIST classification derived from the sum of largest diameters (SLD).
Results or Findings: Median DICE was 0.80. CT lesion tracking was superior to MRI (0.82 vs 0.78). Median absolute error on SLD was 3.2 mm (8.5% of SLD). RECIST classification accuracy reached 71% for PR/SD/PD and 79% for PD vs non-PD.
Conclusion: This pipeline enables longitudinal propagation of target-lesion masks on CT and MRI, advancing toward a semi-automatic RECIST workflow with potential productivity gains and reduced variability.
Limitations: Full RECIST requires integration of non-target lesions. Complete-response cases were absent. The lesion tracking accuracy could be improved. Radiologist oversight remains essential to validate and refine segmentation. Prospective studies are needed to confirm clinical utility.
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: D. Lepot¹, C. Chabot¹, G. Duchêne¹, S. Mandava², M. Fung³, J. Poujol⁴, N. Michoux¹, P. Triqueneaux¹, F. Lecouvet¹; ¹Brussels/BE, ²Atlanta, GA/US, ³New York, NY/US, ⁴Buc/FR
Purpose: The mineral bone has traditionally been inaccessible to MRI. Pseudo-CT MRI sequences now allow visualization of the bone mineral structure for applications like oncology. However these pseudo-CT sequences still have limitations. We compared the accuracy of three pseudo-CT MRI sequences (native Zero-Echo-Time (ZTE), Deep Learning (DL)-Chemical Shift Correction (CSC) reconstructed ZTE (ZTE-DL), and gradient-echo Black-Bone (BB)) in detecting osteolytic multiple myeloma (MM) lesions, using CT as reference.
Methods or Background: Newly diagnosed MM patients underwent ZTE and BB sequences of the lumbar spine, pelvis, and proximal femurs within a 3T whole-body MRI study (Signa Premier, GE HealthCare) (prospective trial: NCT05381077). ZTE-DL was obtained by reconstructing ZTE raw data using an algorithm combining AIR™ Recon DL (ARDL) and CSC.

All patients also underwent 18FDG-PET/CT including an optimized CT within a week. Ten bone regions and two scores (lesion presence/absence, lesion number) were evaluated by three readers. Repeatability and reproducibility (Gwet’s AC1/AC2), differences in lesion number, and accuracy (Acc) were assessed by sequence, region, and reader.
Results or Findings: Ten participants were included. Repeatability was moderate for ZTE (AC1≥0.45), good for ZTE-DL and BB (AC1≥0.60) and very good for CT (AC1≥0.80). Reproducibility was fair for ZTE and BB (AC2≥0.20), good for ZTE-DL (AC2≥0.60) and very good for CT (AC2≥0.80). AccZTE-DL ranged from 80 to 93% with an increase in accuracy ranging from +23% to +32% compared to AccZTE, and equal to +19% compared to AccBB. ZTE-DL detected more lesions than ZTE/BB (+30%/+25%, respectively).
Conclusion: DL and CSC reconstruction improves repeatability, reproducibility, and accuracy of ZTE sequences for detecting MM lesions. ZTE-DL detects more lesions than ZTE/BB.
Limitations: The limitation of the study were a mononcentric/ small cohort and a study only covering the lumbar spine, pelvis, proximal femurs.
Funding for this study: Frederic E. Lecouvet’s research works are funded by the Peterbroeck Research Professorship fund and Allard-Janssen Fund for Cancer Research, Belgian nonprofit organizations.
Caroline Chabot is a PhD student supported by the Fondation Saint Luc, a Belgian non-profit organization.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Nct05381077

Author Block: D. Skwierawska¹, S. Heidarikahkesh¹, D. Bounias², R. J. Jóźwiak³, D. Hadler¹, M. Bachl¹, M. Uder¹, F. B. Laun¹, S. Bickelhaupt¹; ¹Erlangen/DE, ²Heidelberg/DE, ³Warsaw/PL
Purpose: To evaluate the feasibility of automated detection and segmentation of incidental findings in prostate MRI.
Methods or Background: This IRB-approved, retrospective study included n=425 prostate MRI examinations (1.5T and 3.0T), comprising n=306 internal cases from our institution and n=119 external cases from three independent datasets. Manual segmentations were performed for sigmoid diverticulosis (SD), perirectal lymph nodes (PLN), urinary bladder diverticula (UBD), bladder wall thickenings (BWT), inguinal hernias (IH), degenerative changes of the hip (DC), synovial cysts (SC) and hydrocele testis (HT) on T2-weighted images for n=265 internal cases (n=520 ROIs). An nnU-Net model was trained on n=213 of these cases, and the remaining independent n=52 cases were used for the quantitative evaluation of model performance. Further, n =160 additional examinations (n=41 internal, n=119 from three independent external datasets) were evaluated by a radiologist.
Results or Findings: Segmentation performance varied between the incidental findings. Quantitatively, the highest mean Dice scores were achieved for SD (0.80 ± 0.14), HT (0.76 ± 0.20), and DC (0.70 ± 0.07). Radiologist evaluation across four datasets (one internal and three external) demonstrated accuracies of 0.98/0.77/0.93/0.93 for PLN, 0.98/0.87/0.85/0.88 for SD, 0.93/0.90/0.80/0.88 for IH, and 0.93/0.85/0.76/0.80 for DC, with accuracies for most other incidental findings also exceeding 0.85.
Conclusion: Our method can automatically detect and segment incidental findings in prostate MRI across four independent datasets, demonstrating potential to enhance the efficiency and consistency of reporting, supporting further research with larger, more diverse datasets including additional annotations and targets.
Limitations: The dataset was curated to maximise all targeted findings, and despite reflecting real-world prevalence, it remained imbalanced. Rarer and more subjective findings, such as mild bladder wall thickenings or degenerative changes of the hip, were challenging to annotate due to subtle appearances and patient-specific variations.
Funding for this study: Funding from the Bavarian Academic Center for Central, Eastern, and Southeastern Europe (BAYHOST) Scholarship, and from the German Research Foundation (DFG; project 500397400) is gratefully acknowledged.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Name of the ethics committee: Ethics committee of the Friedrich-Alexander-Universität Erlangen – Nürnberg, Medizinische Fakultät
Approval Code: 25-67-Br
Approval Date: 25.03.2025

Author Block: L. T. Rotkopf¹, M. Rempe², C. Strack², H-P. Schlemmer¹, J. Kleesiek²; ¹Heidelberg/DE, ²Essen/DE
Purpose: Medical image segmentation models predominantly operate in the image domain and require fully reconstructed images as input. We investigate a novel deep learning model that directly predicts the k-space representation of breast cancer segmentations from DCE-MRI data and compare its performance to a state-of-the-art image-domain framework.
Methods or Background: The DUKE subset cohort of the publicly available MAMA-MIA breast MRI dataset was selected for this study, consisting of 291 examinations in patients with confirmed breast cancer. For our proposed method, axial k-space slices were generated from the subtraction DCE-MRI volumes using Fourier transforms. A custom 2D U-Net was trained via 5-fold cross-validation to predict the complex-valued k-space tumor mask directly from subtraction k-space. This was compared against a standard 3D nnU-Net baseline trained on reconstructed 3D volumes. The pooled slice-level Dice similarity coefficient for the foreground tumor class was used as primary evaluation metric.
Results or Findings: The proposed k-space segmentation model achieved a mean Dice score of 0.69 ± 0.02 across the five cross-validation folds. The 3D nnU-Net baseline in the image domain achieved a mean Dice score of 0.62 ± 0.26. These results show direct k-space prediction can achieve performance comparable to established image-domain methods.
Conclusion: Our results demonstrate the potential of performing tumor segmentation directly in the k-space MRI domain, with performance levels competitive to established image-domain baselines. Operating in k-space is a novel approach that could open new avenues for developing integrated reconstruction and segmentation pipelines, and may offer advantages in accelerated, undersampled MRI acquisitions.
Limitations: This study was conducted retrospectively on a single-center cohort, which may limit the generalizability of the findings.
Funding for this study: This study was supported by the Jöster Foundation.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: M. Neitzel, M. Sayed, B. Wichtlhuber, E. Frodl, J. Dietz, T. Vogl, P. Reschke, A. M. Bucher; Frankfurt/DE
Purpose: Unreliable AI segmentations can compromise clinical decisions, so we developed and validated Seg-Eval, a model-agnostic, ground-truth-free pipeline that estimates segmentation quality from uncertainty maps of ROI masks on clinical CT.
Methods or Background: We conducted a retrospective multi-cohort study (GastricBCA n=161, PDAC Surgical n=362, PDAC Palliative n=309; total 832 CTs). For each scan, 10 random augmentations (rigid, elastic, or combined) were generated and segmented with TotalSegmentator into 117 ROIs. Masks were mapped back to native space, yielding 8,320 inverse-segmentation masks (ISMs), with the native TotalSegmentator masks on original scans serving as reference. We computed ROI-wise DICE (ISMs vs. native masks), ASSD, coefficient of variation of DICE, augmentation loss, uncertainty (1 − modal agreement), and Pearson correlations; a subset of ROIs (n=90) underwent reader scoring and correction to link uncertainty with DICE improvement.
Results or Findings: Across all cohorts, we observed consistent inverse associations between segmentation quality and variability. DICE correlated strongly and negatively with uncertainty (Gastric r=−0.76; Surgical r=−0.59; Palliative r=−0.57; all p<0.001), indicating that lower performance was accompanied by higher predictive uncertainty. Similarly, higher augmentation loss was linked to increased uncertainty (r=0.70, 0.60, 0.42; all p<0.001), while higher DICE scores were consistently associated with lower variability (r=−0.93, −0.90, −0.91; all p<0.001). Pooling data across cohorts revealed a strong negative correlation between average uncertainty and DICE (Pearson r=−0.69, p<0.001), underlining robustness across subgroups. Reader analysis confirmed that higher-uncertainty ROIs yielded larger DICE gains after correction (r=−0.86, p<0.001).
Conclusion: Seg-Eval provides ground-truth-free, model-agnostic QC by flagging uncertain ROIs and passing reliable results, reducing workload and error propagation while supporting standardised, vendor-neutral integration into clinical workflows.
Limitations: The limitation of the study is that only one segmentation engine was tested.
Funding for this study: Funding was provided by the German Federal Ministry of Education and Research through the RACOON project (reference number 01KX2021).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics approval was obtained by University Medicine Frankfurt (Reference 274/18)

Author Block: J. C. Holzschuh, F. Wäscher, F. Winneknecht, C. H. Ziener, J. Hassel, H-P. Schlemmer, L. T. Rotkopf; Heidelberg/DE
Purpose: Response evaluation in oncology trials relies on manual annotations and therefore limits trial sizes. Because free‑text radiology reports are routinely generated, we hypothesize that large language models (LLMs) can automatically convert them into accurate, structured assessments. However, reliance on cloud‑based or closed models raises privacy concerns. We therefore evaluate local, on‑premises LLMs to accurately extract RECIST response categories from radiology reports.
Methods or Background: In this retrospective study we included 63 patients from clinical trials evaluating immunotherapy for metastatic melanoma. Two state-of-the-art LLMs (MicrosoftPhi-4-14B and DeepSeekR1-32B) were prompted to predict the RECIST category at the first follow-up imaging from the unstructured baseline and follow-up clinical radiology reports. Ground-truth targets were validated clinical RECIST assessments. Performance was measured using multiclass accuracy and sensitivity and specificity for each response category.
Results or Findings: Overall multiclass accuracies for RECIST 1.1 prediction at the first follow-up were 0.56 and 0.64 for Microsoft Phi-4 and DeepSeek-R1, respectively. Class-wise accuracies were 0.55 vs 0.40 for PR, 0.00 vs 0.33 for CR, 0.30 vs 0.55 for SD, and 0.90 vs 1.00 for PD. For predicting PD, Phi-4 and DeepSeek-R1 achieved sensitivities of 0.90 and 1.00, respectively, and specificities of 0.65 and 0.70. No significant differences between models were found (all p>0.05).
Conclusion: On-premises, privacy-preserving LLMs demonstrate promising capabilities in extracting RECIST response categories from free-text radiology reports, especially for differentiating progressive from non-progressive disease. This automated approach potentially reduces the burden of manual evaluation in future clinical trials.
Limitations: Limited include the relatively small sample size, single-institution data, and focus on metastatic melanoma.
Funding for this study: No funding.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Heidelberg

Author Block: J. C. Holzschuh¹, L. Rotkopf¹, F. Bauer², C. Sachpekidis¹, A. Dimitrakopoulou-Strauss¹, H. Goldschmidt¹, M-S. Raab¹, H-P. Schlemmer¹, M. Wennmann¹; ¹Heidelberg/DE, ²Cologne/DE
Purpose: Imaging-derived body composition metrics are increasingly recognized as prognostic biomarkers in oncology. Advances in deep learning now allow automated segmentation of skeletal muscle and adipose tissue, enabling large-scale and potentially routine clinical application. However, the reproducibility of these measures across imaging modalities, particularly between computed tomography (CT) and magnetic resonance imaging (MRI), remains uncertain.
Methods or Background: We analysed same-day whole-body PET/CT and T1-weighted MRI scans from 42 patients from the prospective GMMG-HD7- and GMMG-HD8-/DSMM-XIX-study. A fully automated deep learning pipeline based on the TotalSegmentator Framework was created to segment skeletal muscle, visceral and subcutaneous adipose tissue to derive volume measurements. Volumetric agreement between CT- and MRI-based metrics was assessed using pearson correlation coefficient (r), intraclass correlation coefficients (ICC), and Bland-Altman analyses.
Results or Findings: CT- and MRI-derived body composition measures demonstrated strong correlations (r > 0.97) and high reproducibility for volumetric metrics (ICC > 0.95) across the analyzed compartments, although systematic modality-specific volumetric differences were observed. Compared with CT, MRI yielded lower estimates of adipose tissue and skeletal muscle volumes as Bland-Altman analysis demonstrated a mean bias of -353 cm3 (95% LOA: -2576 to 1869) for subcutaneous adipose tissue, -211 cm3 (-774 to 351) for visceral abdominal adipose tissue, and -2653 cm3 (-4854 to -452) for skeletal muscle.
Conclusion: Automated deep learning-based body composition analysis is feasible and reproducible across CT and MRI, although systematic modality-specific biases were identified. While these are small relative to the overall variance, they should be accounted for in research and clinical implementation. These findings contribute towards using multimodal, imaging-derived body composition biomarkers, potentially guiding future personalized cancer care.
Limitations: While the standardized single-scanner T1-weighted MRI protocol ensures consistency, future studies are required to validate generalizability across scanners, sequences, and larger cohorts.
Funding for this study: No Funding
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Heidelberg

Author Block: R. Hermawan¹, T. Budianto¹, S. J. S. Gardezi², D. Kumar³, A. Gandhamal³, G. Marcel¹; ¹Jakarta/ID, ²Sharjah/AE, ³Delhi/IN
Purpose: Prostate cancer is one of the most frequently diagnosed cancers in men worldwide, representing about 30% of the male cancer burden. Early diagnosis and treatment is key for long-term and progression-free survival in prostate cancer[1]. Artificial intelligence (AI) techniques leveraging deep learning (DL), are increasingly applied in prostate cancer detection to enhance accuracy and lower costs. This study aims to leverage DL for automatic detection , diagnosis and classification of prostate cancer.
Methods or Background: This retrospective study included 40 patients (mean age: 62 years) diagnosed with prostate cancer at Dharmais Cancer Hospital, Indonesia. PIRADS scores were computed using a deep learning–based assessment application (United Imaging Intelligence®). tumor diameter, volume along with right–left (RL), superior–inferior (SI), and anterior–posterior (AP) diameters, were computed. The AI system automatically classified detected prostate tumors into PIRADS categories based on the obtained scores.
Results or Findings: A strong correlation was found between radiologist- and AI-derived prostate volumes (r = 0.976, p = 2.47 × 10⁻²³). Over 40% of cases were classified as PIRADS 5, 22.5% as PIRADS 4, with the remainder in lower categories. Moreover the category wise PIRADS analysis showed highest agreement for PIRADS 5 (62.5%), moderate for PIRADS 3 (42.9%), minimal for PIRADS 4 (11.1%), and no matches for PIRADS 1–2 . Overall the agreement was more reliable in higher categories and minimal in lower categories.
Conclusion: The study underscores the of AI-powered DL application for prostate cancer detection and PIRADS classification. The results exhibit a strong correlation with radiologist-derived prostate volumes and achieved higher agreement in advanced PIRADS categories, highlighting the utility of AI in supporting radiologists.
Limitations: However, more diverse cohort, along with multi-radiologist evaluations, are required to improve reliability across all PIRADS categories and establish broader clinical usage.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: H. Gan, J. Li, Z. Li, Z. Zhou; Wuhan/CN
Purpose: To explore spatiotemporal features from large-scale multi-center contrast-enhanced computed tomography (CECT) data to construct quantitative spatiotemporal models for tumor response in locally advanced rectal cancer(LARC) patients to neoadjuvant chemoradiotherapy(NCRT).
Methods or Background: This study retrospectively collected CECT and clinical information of 337 pathologically confirmed LARC patients from three centers. Two radiologists performed the annotation of tumors and the mesorectum on dual-phase CECT. The cohort was randomly partitioned into training (70%), validation (20%), and test sets (10%). Radiomic extraction yielded 1,223 features per phase. Pathologic response to CRT was assessed histopathologically and graded using the AJCC tumor regression grade (TRG) system (0–3). Treating TRG as a four-class outcome, we evaluated the performance of the radiomics-based multiclass classifier using a one-vs-rest scheme with macro-averaged ROC AUC. We trained thirteen models to predict tumor, mesorectal involvement. Two input schemes were evaluated: single-phase arterial (A) or venous (V) data and a dual-phase V–A difference design, which leverages multiphasic information for improved detectability. Bayesian optimization was used for hyperparameter tuning.
Results or Findings: Among the total population, TRG0 accounts for 9.7%, TRG1 for 13.3%, TRG2 for 41.5%, and TRG3 for 35.5%. For tumor model, arterial phase achieves the highest AUC, at 0.7844; for mesorectum model, the V-A difference performs best, with an AUC of 0.7527. Across Phases A, V, and V-A difference, the tumor model consistently demonstrates better performance than mesorectum , with respective ΔAUC of 0.057, 0.0384, and 0.0097.
Conclusion: The tumour-mesorectum model can effectively explain the spatiotemporal changes of tumors in LARC after NCRT by integrating features across multiple phases and characteristics of different spatial regions. It provides a quantitative tool for evaluating tumor response and mesorectal involvement, and addressing poor efficacy prediction accuracy.
Limitations: Retrospective study.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The IRB of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology approved this work( TJ-IRB202401065).

Author Block: L. V. Sciacqua, B. Dionisi Ferrera, E. Gioscio, C. Marenghi, F. Badenchini, T. Rancati, N. Nicolai, A. Casale, A. Messina; Milan/IT
Purpose: We investigated whether changes in whole-prostate MRI-derived radiomic-features RFs during active-surveillance are associated with histopathological upgrading at repeat biopsy. Specifically, we tested whether increasing time from diagnosis allows aggressive tumor foci to become more detectable on mp-MRI (predictive delta-RFs). We also examined whether these RFs provide signals at diagnostic MRI and whether baseline radiomics adds predictive value to clinical models based on PSA density and number of positive biopsy cores. Frequent ISUP upgrades outside MRI-visible lesions support whole-prostate radiomics.
Methods or Background: We studied 147 consecutive patients (60 with upgrading) from a prospective single-center active-surveillance cohort, monitored with mpMRI (diagnosis, years 1, 2, 4) and targeted biopsies for PIRADS >2 lesions (years 1, 4). Candidate delta-RFs (between diagnosis and last-available mp-MRI) were identified using multiple-ANOVA, and logistic models were built from MANOVA-selected RFs at diagnosis. RFs (T1, T2, ADC maps) were extracted from diagnostic and last-available MRI using a standardized pipeline to calculate absolute baseline RFs and delta-RFs. Associations of delta-RFs with upgrading were tested with multiple-ANOVA, and logistic models were developed using baseline RFs.
Results or Findings: The number of positive cores (OR=1.8, 1 core vs >1) and PSA density (OR=1.6 per 0.05 ng/ml/cc increase) were linked to upgrading (AUC=0.68). The MANOVA model included six delta-RFs: four T2, one T1, one ADC. ADC (morph_vol_dens_aabb) explained most variance. Baseline morph_vol_dens_aabb was associated with decreased upgrading risk (OR=0.90, AUC=0.65). Adding morph_vol_dens_aabb to the clinical-pathological model improved AUC (=0.73).
Conclusion: Prostate radiomics aspects related to the coherence and compactness of ADC maps predict upgrading, indicating prostates with globally organized diffusion are less likely to harbor undetected higher-grade foci.
Limitations: The limitations of the study are the need for independent external datasets to ensure robustness and further analyses of lesion-level RFs.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: J. Zhuang, C. Yuefan, Y. Xie, Q. Ye; Fuzhou/CN
Purpose: The human epidermal growth factor receptor 2 (HER2) expression status varies across different histological types and staging stages of bladder cancer, making it crucial for prognosis assessment. This study investigates the potential of deep learning radiomics models using ultrasound images reconstructed through super-resolution for HER2 status identification.
Methods or Background: A total of 113 patients with confirmed bladder urothelial carcinoma were retrospectively analyzed and divided into the train and test sets in a ratio of eitht to two. A super-resolution (SR) reconstruction technique based on a generative adversarial network (GAN) was applied to increase the spatial resolution of ultrasound images by two and four times. Radiomic features were extracted from the regions of interest (ROIs). The deep learning features were extracted based on ResNet101, and the principal component analysis (PCA) was used for dimensionality reduction. Least absolute shrinkage and selection operator (LASSO) regression was conducted to further screen the radiomic features, deep learning features, and clinical features. Finally, the screened features were involved for machine learning (ML) modeling.
Results or Findings: Four models were developed, including logistic regression (LR), support vector machine (SVM), k-nearest neighbor (KNN), and light gradient boosting machine (LightGBM). Overall, the combined models based on ultrasound images with four times resolution improvement showed the best performance. Among them, the SVM achieved the highest AUC of 0.994 and 0.929 in the train and test sets, respectively. The decision curve analysis (DCA) suggested its potential for assisting clinical decision-making.
Conclusion: The super-resolution (SR) sampling helps to improve the performance of ultrasound-based prediction models. A machine learning model combining clinical features, radiomics features, and deep learning features shows potential for HER2 detection in urothelial carcinoma of the bladder.
Limitations: No limitations were identified.
Funding for this study: Funding was provided by Medical Foundation of Fujian Health Commission of China (no. CX23A010)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The ethics committee notification can be found under the number 2021KY013 of the Ethics Committee of Fujian Medical University Union Hospital