Patients

This prospective study was approved by our institutional review board, and all participants provided written informed consent. We included patients who underwent preoperative multiparametric MRI for suspected renal mass at our institution between December 2019 and February 2023. The inclusion criteria were as follows: (1) patients with primary renal lesions who had not received any prior treatment; (2) patients who underwent preoperative multiparametric MRI (e.g., diffusion kurtosis imaging and intravoxel incoherent motion) within 1 month prior to surgery; (3) patients who underwent surgery with pathologically confirmed RCC. Figure 1 illustrates the reasons for exclusion of certain cases. Ultimately, 47 patients were included in the study.

Fig. 1

Flow chart of patients’ recruitment

Demographic and Clinicopathological Variables, and Definition of Aggressiveness

Demographic characteristics, including age and gender, were retrieved from the electronic medical record system. Pathological information was obtained from the pathology report to identify the aggressiveness of RCC.

MRI protocol

All MRI examinations were performed using a 3.0 T MRI system (MAGNETOM Prisma; Siemens Medical Solutions, Erlangen, Germany) with a 16-channel phased-array body coil to cover the abdomen. Patients were scanned in a supine position. The routine imaging protocol included axial T1-weighted, axial T2-weighted, axial T2 fat-suppression, and coronal T2-weighted sequences. For diffusion-weighted imaging (DWI), multi-b value DWI was acquired using a single-shot diffusion-weighted echo-planar imaging (SS DW-EPI) sequence with 13 b values (0, 10, 20, 30, 50, 100, 200, 300, 500, 800, 1000, 1500, and 2000 s/mm2) to obtain coronal images. Imaging parameters were as follows: repetition time (TR) 1500.0 ms, echo time (TE) 65.0 ms, field of view (FOV) 300 mm × 300 mm, slice thickness 5.0 mm, slice spacing 0.75 mm, scan matrix 128 × 128, and resolution 1.20 × 1.20 × 5.00 mm3. The acquisition time ranged from 8 to 12 min, depending on the patient’s respiratory rhythm. IVIM and DKI imaging analyses, along with image processing and data augmentation methods, are detailed in the Supplementary Materials.

MRI features

The MRI features of RCC assessed with preoperative multiparametric MRI included tumor size, axial and cranio-caudal location, distance to the sinus fat or collecting system, capsular location, tumor shape, tumor margins, invasion of sinus or perinephric fat, encapsulation, preservation of the normal renal profile, extension into the renal hilum, invasion beyond Gerota’s fascia, tumor calcification, hemorrhage, cystic components, renal venous invasion, inferior vena cava invasion, regional lymph node metastasis, distant metastasis, MR RENAL score, and MR pT classification. These features were collaboratively evaluated by two radiologists (G.D. and F.X.) to ensure diagnostic consistency. All remaining samples were independently assessed by a single physician, with all clinical information kept confidential.

IVIM and DKI analyses

The raw images of the multi-b value DWI sequences were processed in Siemens Syngovia Frontier post-processing workstation using MR Body Diffusion software. Parameters were generated based on mono-exponential, biexponential, and diffusion kurtosis models. Apparent diffusion coefficient (ADC) values were calculated using the equation S(b)/S(0) = exp(− b × ADC), where S(b) is the signal intensity at b value, and S(0) is the signal intensity at b = 0 s/mm2. Small b values (0, 10, 20, 30, 50, 100, 200, 300, 500, 600, and 800 s/mm2) were selected for Eq: S(b)/S(0) = (1 − f) × exp(− b × D) + f × exp(− b × Dp). Large b values (800, 1000, 1500 s/mm2) were chosen for Eq: S(b)/S(0) = exp[ (-b × Dapp) + 1/6 × b2 × Dapp2 × Kapp]. Parameters including true diffusion coefficient (D), pseudo diffusion coefficient (Dp), perfusion fraction (f), mean diffusivity (MD), and mean kurtosis (MK) were calculated accordingly. The group b valves were classified as low (0–100 s/mm2), medium (200–800 s/mm2), and high (1000–1500 s/mm2).

Tumor segmentation and the details of image normalization

A radiologist with 5 years of experience manually delineated tumor regions of interest (ROIs) on coronal combined b value DWI images using the publicly available 3D Slicer software (version 5.2.1). These delineations were reviewed and confirmed by a senior radiologist, as illustrated in Fig. 2.

Fig. 2

The ROI was drawn on coronal combined b values DWI Imaging. A, B A case of a 37-year-old man with lowly aggressive e RCC which was pathologically confirmed. C, D A case of a 78-year-old woman with highly aggressive RCC which was pathologically confirmed

Image segmentation was performed by B.W. using OpenCV (version 4.5.3.56), generating masks to distinguish foreground from background pixels. Foreground extraction was achieved using bitwise AND operations with the function cv2.bitwise_and (). For the analysis, 12 b value images (excluding the 2000s/mm2 b value), along with the DWI, IVIM, DKI models, and parametric maps (ADC, D, D*, f, MD, and MK), were used to outline ROIs. Each foreground region was padded to fit a bounding rectangle, reshaped to dimensions of 32 × 32 pixels, and normalized to the range [0, 1]. Additional normalization was applied using the transforms Normalize from Torchvision (version 0.14.1), scaling the pixel values to a mean and standard deviation of (0.5, 0.5), resulting in a final pixel range of [-1, 1] for improved model stability.

Deep learning using CNNs

The LeNet-5 model was utilized to predict RCC aggressiveness based on single b value images, ADC, IVIM, DKI, their parametric maps, grouped b value images, combined b value images, and b value imaging fused with MRI features, as illustrated in Fig. 3. The LeNet-5 architecture comprised two convolutional layers, two max-pooling layers, and three fully connected layers, with a "Sigmoid" activation function for output.

Fig. 3

The framework of the proposed deep learning network for aggressiveness prediction in RCC

The first two layers of the LeNet-5 model were convolutional layers, applying 3 × 3 filters to the input images. These layers detected low-level features, such as edges and textures, and generated feature maps that were passed through ReLU activation functions to introduce non-linearity. Following each convolutional layer, a max-pooling layer with a 2 × 2 filter was applied to reduce the spatial dimensions of the feature maps, thereby decreasing the computational load and preserving essential features. Pooling also enhanced the model’s robustness to variations in the input images. The final pooling layer connected to three fully connected layers with 128, 32, and 1 neuron, respectively. The first two fully connected layers integrated features extracted from the convolutional layers, while the final fully connected layer computed classification probabilities, distinguishing between high- and low-invasive RCC states using a "Sigmoid" function. Deep features extracted from the final fully connected layer were utilized for RCC aggressiveness prediction.

The model was implemented on the PyTorch platform (Torch 1.13.1 with Python 3.8.8) and trained on a CPU. The binary cross-entropy loss function was used to measure the difference between predicted and actual values, optimized using the Adam optimizer with a learning rate of 1e-3. A batch size of 64 was employed to ensure effective convergence, with a maximum of 200 training iterations. Training was halted when the average loss dropped below 0.1, indicating convergence to the global minimum. To ensure the model’s generalizability and prevent overfitting, five-fold cross-validation was performed. This technique divided the data into five subsets, using four for training and one for testing, and repeated the process five times with different training and testing sets. Model performance was evaluated using the receiver operating characteristic (ROC) curve, with the DeLong test employed to compare the diagnostic performance of models incorporating MRI features, parametric maps, and deep learning outputs.

Statistical analysis

Statistical analysis was conducted to compare demographic and imaging features between high- and low-aggressiveness RCC groups. Depending on data distribution, two independent samples t tests, Mann–Whitney U test, Chi-square test, or Fisher’s exact test were employed. The least absolute shrinkage and selection operator (LASSO) regression algorithm was used to identify MRI features associated with high aggressiveness and to develop fusion models incorporating these features with parametric maps and predictive variables. The training and testing processes were repeated five times using five-fold cross-validation to ensure robustness. The performance of various models in predicting RCC aggressiveness was assessed using the ROC curve analysis. The DeLong test was employed to statistically compare the diagnostic performance of ROC curves across MRI features, parametric maps, and deep learning models. Statistical analysis was performed using SPSS (26.0), Python (3.8.8), and R (4.1.2). P values were calculated using two-sample t test, Mann–Whitney U test, Chi-square test, or Fisher’s exact test as appropriate for each comparison. The DeLong test was used to compare the AUC values between different models. P < 0.05 was considered statistically significant for all tests.

To mitigate the risk of overfitting, we implemented multiple strategies during model training. Specifically, early stopping was employed to halt training when the validation loss ceased improving for a specified number of epochs, thus preventing the model from overfitting to the training data. In addition to early stopping, regularization techniques, including L2 weight regularization, were used to penalize large weights in the network and prevent overfitting. The model architecture was designed with dropout layers applied to the fully connected layers, which randomly dropped a fraction of the neurons during training to further improve generalization. Fivefold cross-validation was performed five times to ensure the robustness of the model’s performance across different data splits. This approach helped assess the consistency of model performance and reduce the likelihood of overfitting to any particular subset of the data.

The LASSO regression was employed to select MRI features associated with RCC aggressiveness. The regularization parameter λ, which controls the strength of the penalty applied to the regression coefficients, was optimized using cross-validation. A λ value of 0.0303 was selected based on its ability to balance bias and variance. Features with absolute values of regression coefficients greater than zero were selected, indicating their importance in predicting RCC aggressiveness. To further refine the model, features were only retained if they showed statistical significance, with a p value threshold of < 0.05. The final selected features—sinus/perinephric fat invasion, tumor stage, and tumor size—were consistently associated with high aggressiveness. The feature selection process was repeated across multiple iterations, ensuring the robustness of the selected features. Features that were consistently chosen across these iterations were included in the final predictive model.