Neuroblastoma (NB) is a type of peripheral neuroblastic tumor commonly found in children and characterized by obvious heterogeneity in biological behavior and rapid development. Determining the differentiation type is helpful in assessing the prognosis of neuroblastoma for making early judgments regarding postoperative treatment options. Whole-slide images (WSIs) of the NB have ultrahigh resolution and contain rich information, facilitating clinical interpretation. However, early diagnosis is time-consuming and poses significant challenges. Considering the complex cellular environment and heterogeneity of NB, this study proposes a novel network, CSA-U-Net, for cell segmentation and classification of NB WSI. Additionally, a cross-pseudo-supervised (CPS) approach, combining different proportions of labeled and unlabeled data, is used for training, which improves the robustness and generalization ability of the model, thereby assisting pathologists in clinical diagnosis, reducing their workload, and decreasing the misdiagnosis rate.

To address the cell-level data labeling problem, this study adopt a deep learning method based on CPS, fully utilizing the distributional characteristics of unlabeled data and combining a small amount of labeled data, to improve the model’s generalization ability by having the two branches supervise each other. To address the complex cellular environment and heterogeneity of NB, channel and spatial attention modules are added to the bottleneck layer of U-Net network. The proposed novel network, CSA-U-Net, is served as the base network for the CPS model, effectively improving model accuracy. Finally, the K-means algorithm is used to classify and count poorly differentiated and differentiated NB cells in the pathology slide images. The percentage of differentiated NB to the total number of tumor cells is calculated, to assist pathologists in determining histopathological typing.

The CPS approach for NB WSI segmentation is shown in Fig. 1, with CSA-U-Net as the underlying network for the two branches (Fig. 5). The CSA-U-Net network was compared with U-Net, DeepLabv3+, PSPNet, HrNet, SA-U-Net, HoVer-Net, and MEDIAR. The results showed that the CSA-U-Net outperforms the other methods in all indicators. The F1 score was 79.05% in poorly differentiated cells and 62.21% in differentiated cells, and the accuracy was 96.78%, which is an improvement compared with that of the traditional U-Net (Table 1). In the prediction result graph, the prediction results of CSA-U-Net exhibit more accuracy, clearer boundaries, and less noise in the image, relative to other networks. A lower error rate is observed in the regions prone to erroneous segmentation (Fig. 8). Next, the performance difference of the CPS method with CSA-U-Net as the base network, was explored for labeled to unlabeled data ratios of 1∶1, 1∶2, 1∶3, and 1∶4. The results show that the segmentation accuracy of the model gradually increases with an increase in the amount of unlabeled data, and the F1 score of the model improves faster before the ratio of labeled to unlabeled data reaches 1∶3. After the ratio reaches 1∶4, the model enhancement is slower, and the speed of accuracy enhancement decreases significantly (Table 2). Subsequently, the CPS method was compared with other semi-supervised methods, at a 1∶3 ratio of labeled to unlabeled data. The CPS method showed the best detection performance, with F1 score of 80.99% in poorly differentiated cells, 65.40% in differentiated cells, and 97.99% accuracy (Table 3). Finally, the different types of cells in the prediction results were counted using the k-means method and compared with the gold standard of physicians (Fig. 9). The average accuracy of the counting results of poorly differentiated and differentiated NB cells was 94.00% and 89.89%, respectively (Table 4). This result indicates that the method in this study excels in the counting accuracy of poorly differentiated and differentiated cells and operates stably in images of any size, further validating the reliability of the method.

To address the problem of large amounts of cellular data and heavy labeling in NB images, this study adopted a CPS approach for model training. By introducing unlabeled data during training, the model can better capture the features of poorly differentiated and differentiated cells, thereby more accurately extracting and categorizing these cells from tissue backgrounds and displaying better adaptation to the variability and complexity of different samples. The CPS approach ensures the consistency of the two branches in terms of network structure while making them differ in parameter space through different initializations and independent training, which drives the model to learn a more robust and comprehensive feature representation. Meanwhile, for the features of NB pathology slide images, this study proposes a CSA-U-Net network model, incorporating an attention mechanism based on the original U-Net network, which further improves the accuracies of the segmentation and classification results. This study is based on the CSA-U-Net network and effectively integrates labeled and unlabeled data using a CPS semi-supervised model. The experimental results show that the CSA-U-Net network exhibits better performance on the NB dataset than existing control methods, with the segmentation accuracy of the model gradually improving as the amount of unlabeled data increases, which further validates the effectiveness of the CPS method. Finally, the K-means method was used to count the different types of cells in the model prediction results for pathological staging. The method proposed in this study effectively reduced the workload of pathologists, improved diagnostic efficiency, and is of great significance in determining the prognosis of NB.