Quality control for images and annotationsIn this study, rigorous quality control is applied to MRI images and annotations. Firstly, to ensure that the population of research subjects is sufficiently consistent on key characteristics, all MRI images are selected based on uniform inclusion and exclusion criteria. Secondly, each image underwent quality assessment to ensure the absence of motion blur or artifacts, and to maintain sufficient clarity for accurately depicting details of the regions of interest. For image annotations, experienced radiologists are tasked with precise tumor localization and annotation. To guarantee the accuracy and consistency of annotations, a double-review process is employed, where one radiologist performs the annotation and another experienced radiologist reassess each annotation to ensure reliability. We calculate intra-rater reliability using the Dice similarity coefficient, which indicates if the same voxels are being selected as part of the lesion mask or not. For Dice calculation, we compare the annotations of two radiologists for all 275 cases, and the intra-rater Dice coefficient is 0.870. We also calculated the intraclass correlation coefficient (ICC) for the lesion volumes. The ICC ranges from 0–1; 1 is total agreement. The intra-rater ICC is 0.988. These quality control measures aim to enhance the validity and credibility of the dataset in bladder cancer diagnosis research.Experimental verification in federated learning tasksTo assess the enhancement of accuracy and generalization provided by FL, we utilize FL methods, centralized training (mixed data from four centers), and single-center training to develop a MIBC prediction model or automated tumor segmentation model, respectively. To build the baseline of FL in the dataset, we conduct a survey on classical FL methods.These methods include FedAvg6, SiloBN17, FedProx18, and FedBN19, each with distinct algorithm designs and implementation details. FedAvg6 is a foundational algorithm that trains a global model across multiple clients while keeping data localized. FedAvg involves initializing a global model, performing local training on each client, sending model updates to the server, and averaging these updates to form a new global model, iterating until convergence. SiloBN17 addresses data heterogeneity in multi-center medical investigations by combining a local batch normalization (BN) layer with center-specific statistics. This approach results in a model that is jointly trained and tailored to each center. SiloBN enhances robustness under varying data conditions while minimizing the risk of information leakage by avoiding the sharing of center-specific activation statistics. FedProx18 improves the handling of non-IID data through a re-parameterization module and targeted parameter modifications for individual clients. FedProx also allows for varying quantities of local tasks across devices and stabilizes the method with an approximation term. FedBN19 facilitates feature transfer among heterogeneous clients by enabling the exchange of extracted model attributes instead of raw data. Local BN is employed to align feature distributions across clients, ensuring consistency and supporting local model training.We use these four FL methods to build corresponding baseline of the dataset in diagnosing MIBC and performing automatic segmentation of BCa. Subsequently, we compare the performance of these methods on the test set (Tables 3 & 4).Table 3 Results of Classification Task for the Dataset.Table 4 Results of segmentation tasks for the Dataset.In this study, we conduct all experiments using PyTorch for training on NVIDIA A100 GPUs. The models are trained in a Python environment (version 3.8; https://www.python.org/), utilizing PyTorch (version 1.13.1; https://pytorch.org/). Our computing system is equipped with Intel Xeon Gold 6326 processors.We refine the preprocessing of bladder MR images, adapting to different tasks in this study. Each slice of the 3D T2WI is cropped to uniform dimensions. For classification task, original T2WI slices are cropped to create 128 × 128 patches centered around the tumor annotations. For segmentation tasks, the cropping frame size of T2WI slices is set at 160 × 160. The cropped frame, centered around annotations, is randomly offset by 10 to 15 pixels in the x-y axes. Figure 5 shows an overview of the experimental process.Fig. 5An overview of the experimental procedure. Each center acts as a client. For each round of communication, a certain percentage of clients are randomly selected to the train local model and send the local model to the server. The server aggregates the new global model and updates the model of client.We use image augmentation techniques, including horizontal and vertical flipping, image cropping, and affine transformations, to optimize the utilization of our data representation. For model optimization, we utilize the Adam optimizer with a fixed learning rate of 1e-05. In model training, the Cross-entropy loss20 function is adopted for classification tasks, while Dice loss21 is utilized for segmentation tasks. The batch size is set to 24, and the training is conducted over 500 epochs.Considering the limited sample size from center 2, 3, and 4, we select U-Net22 network, which is effective with small datasets, as the backbone for our segmentation tasks. We select ResNet-5023, a well-regarded classification network, as the backbone for the classification tasks. We randomly select 40% of data from each center for testing in classification tasks. For segmentation tasks, a randomly selected subset of 30% patients from each center is used to assess the performance of the model.To balance computational efficiency and model accuracy, we set the proportion of clients participating in federated aggregation per round is set to 0.5, meaning approximately half of the clients participate in each global model aggregation. The number of local training epochs before each aggregation is set to 1, indicating that the local model trains for one epoch before aggregation. The batch size for local model training is set to 24. In this study, we utilize the Area Under the ROC Curve (AUC) to evaluate the performance of the classification models, and Dice similarity coefficient (DSC) to evaluate the segmentation performance.The classification task results for the dataset are presented in Table 3. The Centralized training, which combines the training data of four centers, exhibits the highest AUC, with a mean value of 0.866. Among FL methods, SiloBN achieves the highest average AUC (0.849), followed by FedBN (AUC = 0.842). FedAvg and FedProx show competitive performance with AUCs of 0.839 and 0.824, respectively.The prediction model trained on a single center demonstrates average AUCs ranging from 0.783 to 0.811. Among these, the model trained by Center 1 achieves the highest diagnostic accuracy. Notably, the diagnostic accuracy of all models trained on a single center is lower than the FL method.Centralized training achieves the highest automatic segmentation accuracy (DSC = 0.841), as detailed in Table 4. The model trained by the data from Center 1 achieves the highest single-center training results (DSC = 0.770), which may be due to its larger data volume. All the four FL methods outperform single-center training. Among them, the FedProx method achieves a segmentation accuracy (DSC = 0.840) second only to centralized training. FedBN and SiloBN show competitive performance with DSCs of 0.837 and 0.831, respectively. It is noteworthy that the FL methods not only achieve superior segmentation accuracy over single-center training on average DSC, but this trend is consistently observed across each center. Figure 6 presents the segmentation results of four typical cases of the dataset with different methods, indicating that the models trained by centralized training and FL are more accurate in segmentation.Fig. 6Four typical cases from the dataset. Each case includes the T2-weighted image, segmentation annotations (ground truth), and the predicted segmentation results.It is worth noting that models trained at a single center do not always perform well on test data from their own center, both in Classification and Segmentation tasks. The analysis of the data reveals several reasons for this issue. Firstly, each center’s dataset may not capture the full range of variability in the overall data distribution, leading to models that are overly specialized and fail to generalize well even within the same center. For example, the model trained at Center 1 has an AUC of 0.720 on its own test data but performs better on data from other centers, achieving an AUC of 0.900 on Center 3’s test data. Secondly, small sample sizes and data noise within each center can affect the model’s ability to learn robust features, leading to suboptimal performance. This is evident in the model trained at Center 4, which has an AUC of 0.750 on its own test data. These performance discrepancies highlight the challenges of single-center training and emphasize the advantages of centralized and federated learning approaches in developing more robust and generalizable models.
