Patients

This retrospective study was approved by the Institutional Review Board of the First Hospital of Shanxi Medical University. Written informed consent was obtained from all patients in this study. We conducted a retrospective search within the pathology and radiology databases of the First Hospital of Shanxi Medical University, spanning from January 2018 to June 2021, to identify patients diagnosed with BCa. Subsequently, we analyzed the clinical data of 229 patients with pathologically confirmed BCa.

Inclusion criteria were defined as follows: (1) confirmation of BCa through pathology following TURBT or RC; (2) initial TURBT or cystectomy procedure; (3) MRI examination conducted within 3 weeks prior to surgery; and (4) the availability of complete clinical and pathological data. Exclusion criteria comprised the following: (1) receipt of preoperative radiotherapy or chemotherapy; (2) cases exhibiting poor image quality or tumors with a diameter less than 3 mm, making delineation unfeasible; (3) instances with incomplete clinical and pathological data; or (4) patients lost to follow-up or follow-up was less than 2 years. The process of case inclusion and screening is presented in Fig. 1.

Fig. 1

Flowchart shows selection criteria for the 229 patients in the study group

Clinical data collection and patient follow-up

We collected clinical data from 229 patients with confirmed BCa, including seven parts: demographic characteristics (such as age at diagnosis, BMI, gender, smoking, and drinking), clinical characteristics (such as frequent urination, urinary urgency, odynuria, urinary incontinence, low back pain, and previous malignancies), serum laboratory information (such as total cholesterol, triglyceride, high-density lipoprotein, low-density lipoprotein, urea, creatinine, and uric acid), tumor characteristics (such as tumor location, tumor size, and tumor number), pathological data (such as pathological grading, and MIS), treatment information (such as infusion drug, and surgical methods), and survival information (RFS).

Following surgical intervention, each patient underwent a meticulously planned follow-up regimen, involving an initial assessment within 3–5 months post-surgery, subsequent evaluations every 6 months for a duration of 2 years, and annual appointments thereafter. These follow-up assessments entailed comprehensive cystoscopy and imaging examinations (CT or MRI) to scrutinize for any signs of suspected bladder tumor recurrence. For the purpose of this study, recurrent tumors were defined as tumors that reappeared within the bladder, prostate, urethra, pelvis, or ileum subsequent to surgical intervention. By the end of follow-up, 81 BCa patients experienced a recurrence. Importantly, we documented the time to RFS for each patient, calculated from the date of their initial surgery.

MRI protocol

All MRI scans were conducted using a 3.0-T MRI scanner (Skyra: Siemens, Erlangen, Germany) equipped with an 8-channel truncal phased-array coil or body coil. Specific scan range, scan sequence, and detailed scan parameters are provided in Additional file 1.

Tumor ROI segmentation

Two experienced radiologists, referred to as Reader1 and Reader2, each possessing over 8 years of experience working with MRI, manually delineated the entire bladder tumor using ITK-SNAP 3.8.0 (http://www.itk-snap.org/). Importantly, they remained blinded to the patient’s pathological findings throughout the procedure. Subsequently, the axial images from T2WI, DWI, ADC, and DCE were segmented to extract the volume of interest (VOI). In cases where disagreements arose, a resolution was reached through consultation.

Additionally, to assess the consistency of the image feature delineation both within and between observers, 1 month later, the VOI of 15 recurrent and 15 non-recurrent patients were randomly selected and delineated by a single radiologist, Reader1. This allowed for the evaluation of the intraclass correlation coefficient (ICC) for the delineated image features.

The tumor delineation contours of all sequences are shown in Fig. 2. The definition of tumor margin on different weighted images needs to be determined based on the signal characteristics. The specific rules for this process can be found in Additional file 2.

Fig. 2

A 69-year-old male patient presented with recurrent BCa on follow-up surveillance and an 85-year-old male patient with follow-up surveillance for non-recurrent BCa

Radiomics feature extraction

We used the open-source software FAE (https://github.com/salan668/FAE) using PyRadiomics package to extract radiomics features from T2WI, DWI, ADC, and DCE images of BCa patient. The VOI delineated on T2WI, DWI, ADC, and DCE images for each BCa patient underwent a series of preprocessing steps. The radiomics feature extraction steps and feature categories are described in Additional file 3.

Development of the radiomics nomogram

Firstly, the reproducibility of each feature delineation process was quantified by evaluating the ICC values of the inter- and intra-group datasets, and the radiomics features with ICC < 0.75 were excluded for subsequent analysis. The remaining stable features were normalized and significant radiomics features were selected using univariate Cox regression with p < 0.05. Univariate analysis and LASSO regression algorithm were used to select the optimal feature subset. The RFS prediction model of BCa was constructed based on the optimal radiomics features, and the radiomics score (radscore) of each patient was calculated. The cutoff value for the high and low-risk groups were identified by the median radscore in the training set, and Kaplan-Meier analysis was used to assess the potential association between radscore and RFS, which was validated in the validation set.

Development of the clinical nomogram

Univariate Cox regression analysis was utilized to assess the association between clinical features and the RFS in patients. After finding a statistically significant difference (p < 0.05) between the recurrent and the non-recurrent, we conducted a multivariate Cox regression analysis and then selected independent predictors of BCa recurrence with p < 0.05. Based on this, a clinical model was developed and used for validation.

Development of the radiomics-clinical nomogram

By incorporating the independent risk factors identified in the clinical model and integrating them with the radscore derived from the radiomics model as covariates, we have devised a practical and clinically relevant radiomics-clinical nomogram. This nomogram serves as a valuable tool for predicting early BCa recurrence and individualized RFS. To evaluate the performance of the nomogram and its goodness of fit, we utilized metrics such as the C-index and calibration curve, both in the training set and the validation set. In order to evaluate the nomogram’s diagnostic capabilities, we employed the net reclassification improvement (NRI) and compared it against the radiomics model and clinical model. Lastly, we conducted decision curve analysis (DCA) for all three models, providing insights into the clinical validity and utility of our proposed tool.

Statistical analysis

The statistical analyses for this study were conducted using R version 4.2.0 (https://www.r-project.org/) and SPSS 26.0 (http://www.spss.com.cn). The one-sample Shapiro-Wilk test was used to assess the normality of numerical variables. For normally distributed data, we presented results as mean ± standard deviation (M ± SD), while non-normally distributed data were represented as median (interquartile range (IQR), 25th and 75th percentiles). Two-sample t-test was used to compare normally distributed data between groups, and the Mann-Whitney U test was used to compare non-normally distributed data. The chi-square test was used to analyze categorical data.

To gauge the relationship between the radscore derived from the prediction model and the RFS status of patients, we employed Kaplan-Meier analysis and Cox regression analysis to compute early recurrence rates. Statistical significance was established at a threshold of p < 0.05. The C-index was utilized to evaluate the model’s accuracy in predicting recurrence stratification and RFS performance in two sets.