This section systematically presents a 2D segmentation method for renal tumors based on CT images. First, we provide a detailed description of the construction of the experimental dataset and the standardized preprocessing pipeline. Subsequently, we comprehensively describe the design and implementation of the DeepLabV3+ framework along with its novel modules.

In this study, we conducted a retrospective analysis using a private dataset obtained from the Lanzhou University Second Hospital. The dataset comprises 3D CT renal tumor images from 218 patients who were treated between 2016 and 2022. Data collection for this study was conducted over the 6-month period beginning on October 28, 2022. Notably, the dataset used in this study does not include any personal identifying information of the patients. The dataset is provided in Neuroimaging Informatics Technology Initiative (NIfTI) format with corresponding segmentation labels. In addition to CT imaging data, it encompasses various histological subtypes of renal tumors, including ccRCC, papillary renal cell carcinoma (pRCC), and chromophobe renal cell carcinoma (chRCC). The tumors were precisely classified into stages I, II, III, and IV based on the international TNM staging system. Moreover, the dataset contains fundamental clinical attributes such as patient age and gender; however, these clinical variables were collected but not used in the current segmentation algorithm, which operates solely on 2D CT images. The segmentation annotations were manually performed by experienced radiologists based on surgical pathology results to guarantee accurate tumor localization. Nontumor cystic lesions were strictly excluded to enhance the dataset’s quality. The annotations include 2 categories: background and renal tumor, as illustrated in Figure 2A.

Additionally, experiments were conducted on the publicly available KiTS19 dataset from the 2019 Kidney and Kidney Tumor Segmentation Challenge [33]. This dataset comprises original abdominal CT images and manually annotated label images by physicians, with annotation classes for background, kidney, and renal tumor, as illustrated in Figure 2B. To enhance data complexity and diversity, and to create a more challenging training environment for the network, we incorporated diverse cases from 210 patients. Notably, this dataset is publicly available on GitHub [34] and has been approved by the Institutional Review Board of the University of Minnesota. Throughout this study, we strictly complied with the terms of the KiTS19 license [33].

The private dataset and the public KiTS19 dataset exhibit complementary characteristics in terms of data construction, collectively offering diverse perspectives for evaluating model performance. The private dataset was meticulously annotated on a per-case basis by experienced radiologists, and a single representative slice was extracted for each patient using dedicated software from 3D volumetric NIfTI-format data. In contrast, the KiTS19 dataset was preprocessed using the official pipeline, generating a series of continuous slices from the raw volumetric scans. To comprehensively assess the generalizability of the proposed model, we conducted experiments on both datasets, thereby validating its robustness and adaptability across different data acquisition and annotation protocols.

This study was approved by the Ethics Committee of the Second Hospital of Lanzhou University (approval number: 2022A-575), ensuring compliance with medical ethical standards. Given the retrospective nature of the study and the use of deidentified data, the requirement for informed consent was waived by the ethics committee. All imaging data were fully anonymized prior to analysis, and no direct or indirect patient identifiers were accessible to the research team. No financial or material compensation was provided to participants for their inclusion in this study. No images included in the paper or supplementary materials contain identifiable information of individual participants.

Fundamentally, image segmentation can be viewed as a pixel-level classification problem, where each pixel is classified as either background or target. For segmentation tasks, it is ideal for the network to correctly predict target pixels as true positives (TP) and background pixels as true negatives (TN), whereas misclassifying background pixels as target or vice versa results in false positives (FP) and false negatives (FN), respectively. In this study, the performance of the proposed GAM-DeepLabV3+ network in renal tumor segmentation is quantitatively evaluated using 6 metrics: intersection over union (IoU), Dice, recall, precision, accuracy, and 95% Hausdorff distance (HD95). IoU is a widely used metric in image segmentation that calculates the ratio of the intersection to the union of the predicted segmentation and the ground truth, as given by the following formula:

The Dice coefficient, which measures the similarity between 2 samples on a scale from 0 to 1 (with values closer to 1 indicating better performance), is calculated as follows:

Recall is defined as the proportion of TP pixels correctly identified by the model relative to all actual positive pixels in the ground truth, computed as:

Precision refers to the ratio of TP pixels among all pixels predicted as positive by the model, computed as:

Accuracy denotes the proportion of all pixels that are correctly classified by the model, measuring the overall prediction accuracy with the formula:

HD95 refers to the 95th percentile Hausdorff distance between the predicted segmentation boundary and the ground-truth boundary. It is less sensitive to extreme outliers than the maximum Hausdorff distance and was used to evaluate boundary accuracy. It is calculated as follows:

where d(A,B)={min{‖b−a‖2:a∈A}} and d(B,A)={min{‖b−a‖2:b∈B}}. Here, A and B denote the boundary point sets of the predicted segmentation and ground-truth mask, respectively. In this study, HD95 was reported in pixels.

Our proposed loss function is a combination of binary cross-entropy (BCE) loss and Dice loss, referred to as BCEDiceLoss. BCE loss, a special case of cross-entropy loss suitable for binary classification problems, is computed as follows:

Here, y represents the ground truth label (0 or 1), and p denotes the model’s predicted probability. The BCE loss function quantifies the discrepancy between the ground truth and the predicted probabilities.

By combining BCE loss with Dice loss, both classification accuracy and segmentation detail can be simultaneously optimized during training. Specifically, the total loss can be computed using the following formula:

Here, BCE refers to the binary cross-entropy loss, and Dice denotes the Dice coefficient (computed with a smoothing term), so that the term (1 − Dice) constitutes the Dice loss and a higher Dice coefficient yields a lower loss. λ is a hyperparameter used to balance the contributions of the 2 loss terms, which was empirically set to 2 in all experiments. During training, the model optimizes both BCE and Dice objectives concurrently, thereby improving its overall performance.

By integrating the metrics, we can comprehensively and objectively evaluate the model’s performance under various conditions, thereby providing robust data support for subsequent model optimization and clinical applications.

In our experiments, we used the CT data described in the data preparation section and established the deep learning environment using Python 3.8 (Python Software Foundation) and PyTorch 1.11.0 (PyTorch Foundation, The Linux Foundation) on Ubuntu 20.04 (Canonical Ltd). The hardware configuration included an AMD EPYC 9754 processor and an RTX 4090D GPU with 24 GB of memory. During training, the batch size was set to 8 and the maximum number of epochs was set to 200, with validation performed at the end of each epoch. An early stopping strategy was used to prevent overfitting. The Adam optimizer was adopted with an initial learning rate of 1×10^–4, a weight decay of 1×10^–4, β1 of .9, and a minimum learning rate of 1×10^–5. To ensure reproducibility and assess result stability, all compared models were trained and evaluated 5 times using different random seeds (42, 123, 456, 789, and 1024), and the results are reported as mean (SD). All dataset splits were performed at the patient level to prevent data leakage. For the private dataset, 218 patients were divided into 153 training, 43 validation, and 22 testing cases, with one representative slice extracted per patient. For the KiTS19 dataset, 210 patients were divided into 147 training, 42 validation, and 21 testing cases; each patient contributed multiple continuous slices containing tumor regions, resulting in 4800 slices for training and evaluation. All backbone networks used in the ablation experiments were initialized with ImageNet pretrained weights. For the no-new-Net (nnU-Net) baseline, we adopted a 2D nnU-Net implementation and trained it under the same unified experimental protocol as the other baseline models. The complete self-configuring nnU-Net pipeline, including automated experiment planning, preprocessing, configuration selection, and postprocessing, was not used. To assess statistical significance, paired t tests were conducted across the 5 independent runs for Dice scores between the proposed method and baseline models, with statistical significance defined as P<.001.

During the evaluation of the proposed GAM-DeepLabV3+ model, training and testing were performed on both a private dataset and the public KiTS19 dataset. All splits were performed at the patient level, with 70% for training, 20% for validation, and 10% for testing, with identical hyperparameter settings and training strategies used to objectively compare the model’s performance across different data sources. The proposed model converged stably within 200 epochs on both datasets, with no signs of overfitting observed under the patient-level experimental setup. Figure 7 presents the final segmentation visualization results.

We randomly selected 2 groups of CT image predictions from the test sets of both the private dataset and the KiTS19 dataset for comparative analysis, as shown in Figure 7. This visual comparison demonstrates the segmentation results of different network models on renal tumors, with the white regions representing the tumor segmentation. It can be observed that traditional classic networks such as U-Net and feature pyramid network (FPN) produce relatively coarse segmentation boundaries for complex-shaped and small-volume tumor targets, often leading to segmentation errors due to insufficient deep feature extraction, which results in the failure to capture certain detailed regions. In contrast, the proposed GAM-DeepLabV3+ network performs well in segmenting small and irregular lesion areas. Compared with other classical networks, it reduces both undersegmentation and oversegmentation issues, primarily because the network is able to filter channels that contain critical information through dual attention modules, thereby reducing feature loss and enhancing overall accuracy. Consequently, it can, to a certain extent, resolve the separation between small tumor regions and the kidney, and compensate for missegmentation or omission errors observed in other networks, thereby providing valuable assistance for clinical diagnosis of renal tumors.

This study introduced GAM-DeepLabV3+, a lightweight, automated framework for renal tumor segmentation, and validated its efficacy on both a single-center clinical dataset and the multicenter KiTS19 benchmark. Our primary finding is that the proposed model achieves superior segmentation precision while maintaining high computational efficiency. On the private dataset, the model yielded a mean DSC of 0.939 (SD 0.008) and a mean HD95 of 1.485 (SD 0.522) pixels, consistently outperforming baseline architectures such as the standard DeepLabV3+ (DSC: mean 0.911, SD 0.008). The robust performance on the KiTS19 external validation set (DSC: mean 0.928, SD 0.006; HD95: mean 2.536, SD 0.742 pixels) further demonstrates the framework’s strong generalizability across heterogeneous data acquisition protocols and varying annotation standards.

Ablation analysis systematically quantified the contributions of the individual architectural innovations (Table 4). Transitioning from an Xception to an enhanced MobileNetV2 backbone provided a 0.5% increase in DSC (from 0.917 to 0.922), confirming that a lightweight backbone can preserve, and even refine, feature representation while significantly reducing computational complexity. Most notably, the substitution of CBAM with the GAM resulted in the most substantial gains, improving the DSC by 1.1% over the CBAM-integrated version. This suggests that the sequential channel-spatial attention design of GAM contributes to suppressing background noise and refining boundary delineation; however, because our evaluation did not include a tumor size–stratified analysis, a specific benefit for small renal lesions remains to be confirmed quantitatively. Furthermore, the reduction in per-slice inference latency (from 21.8 to 14.2 ms/slice on the private dataset, an approximately 1.5× speed-up) indicates improved computational efficiency, although real-time clinical use would require further validation.

The performance metrics in Table 5 illustrate the advantages of the proposed approach relative to previously reported methods on the KiTS19 dataset. Early CNN-based approaches provided a foundation for automated renal tumor segmentation but faced inherent limitations in handling morphological complexity. Yang et al [42] achieved a Dice of 82.6% using a CNN architecture with a 3-stage augmentation strategy; while this result demonstrated the promise of deep learning for this task, the approach encountered difficulties with structurally complex tumor presentations. da Cruz et al [43] advanced the field by introducing a 2.5D DeepLabV3+ model with cross-plane feature fusion, reaching a Dice of 85.17%. Ji et al [44] introduced ASD-Net, a U-Net variant incorporating asymmetric spatial-channel convolution modules, achieving a Dice of 85.22%. In the same period, Ruan et al [45] proposed MB-FSGAN, a multiscale bipath generative adversarial network that reached 85.9%, although its feature extraction remained less sensitive to small lesions.

The growing interest in 3D architectures subsequently opened new directions for performance improvement. Türk et al [46] combined residual connections with dense block structures in a hybrid V-Net, achieving a Dice of 86.5%; however, this came at the cost of substantially increased model parameters and computational overhead. Sun et al [47] balanced computational cost against accuracy through a multiscale feature fusion strategy in their 2.5D MFFAU-Net, reaching 87.5%. Jackson et al [48] demonstrated that a 3D CNN augmented with data augmentation could attain a Dice of 88.5%, though the shallow architecture constrained the model’s capacity to represent heterogeneous small-tumor features. More recently, attention mechanisms have driven further performance gains: Sun et al [49] proposed FR2PAttU-Net with a dual-path feature refinement module and pyramid attention unit, achieving a Dice of 91.1%, albeit with a 32% increase in parameter count attributable to its cascaded module design. Notably, the external scores in Table 5 are tumor-class Dice from the original 3-class KiTS19 task (background, kidney, and tumor), whereas our model adopts a binary tumor-versus-background formulation with the kidney merged into the background; the comparison is thus indicative rather than strictly controlled. Moreover, merging the kidney into the background does not necessarily simplify tumor delineation, as the renal parenchyma is the tissue most readily confused with tumor.

The proposed GAM-DeepLabV3+ framework achieved a Dice of 92.8% on the KiTS19 test set, which is numerically higher than the 91.1% Dice reported for FR2PAttU-Net [49], while maintaining an accuracy of 99.2% and a more favorable computational profile. As noted in the footnote to Table 5, this comparison is indicative rather than a strictly controlled head-to-head result, because the external KiTS19 scores derive from the original 3-class task whereas our model adopts a binary tumor-versus-background formulation. On the private dataset, GAM-DeepLabV3+ outperformed both ASD-Net (Dice: 0.876, IoU: 0.810) and DeepLabV3+2.5D (Dice: 0.888, IoU: 0.806), achieving a mean Dice of 0.939 (SD 0.008) and a mean IoU of 0.928 (SD 0.012; Table 6). These results collectively support the effectiveness of combining global context modeling through the GAM module with a lightweight MobileNetV2 encoder for renal tumor segmentation under both data-rich and data-limited conditions. We note that these 2 comparators were reimplemented in-house and evaluated as single-run reference comparisons on the private test set, rather than under the multiseed protocol used for the other baselines in Table 1; they are therefore reported as single-point values.

An online demonstration platform based on GAM-DeepLabV3+ has been made freely accessible at CPPDD [50] (for demonstration purposes only; not intended for upload of real patient data), as illustrated in Figure 8. The platform is intended for research-context demonstration only and is not validated for clinical use.

This study has several limitations that warrant consideration. The private dataset comprised 218 patients from a single center, with one representative 2D slice extracted per patient. Although this strategy is consistent with the 2D segmentation paradigm adopted here, it discards volumetric continuity information and may not fully reflect the spatial heterogeneity of renal tumors across contiguous axial slices. The relatively high performance metrics observed on the private test set may in part reflect the representative nature of the selected slices rather than the full spectrum of imaging variability encountered in clinical practice. To partially address this, external validation was performed on the KiTS19 dataset, which includes multiple consecutive slices per patient and data from multiple centers, providing a more challenging and diverse evaluation environment. The model achieved a mean HD95 of 1.485 (SD 0.522) pixels on the private test set and a mean HD95 of 2.536 (SD 0.742) pixels on KiTS19, suggesting reasonable spatial consistency of the 2D predictions; nevertheless, formal 3D reconstruction validation—including volumetric Surface Dice and 3D HD95 computed across the full z-axis stack—to confirm the absence of stacking artifacts remains an important direction for future investigation. This constitutes a key limitation of the current work. In addition, because the external KiTS19 results in Table 5 were obtained under different task definitions (3-class vs binary), preprocessing, and evaluation protocols, the numerical differences reflect contextual competitiveness rather than a controlled demonstration of superiority. Finally, although the architecture is motivated in part by the difficulty of small renal tumors, no tumor size–stratified evaluation was performed; the corresponding claims are therefore qualitative, and a volume-stratified analysis of small-lesion performance is left for future work. Furthermore, because only 2D slices containing annotated renal tumors were retained and tumor-free (negative) slices were discarded during dataset construction, the model’s false-positive behavior on normal, tumor-free tissue and its performance on complete patient volumes that include negative slices were not evaluated in this study; assessment on full CT volumes therefore remains an important direction for future work.

Although nnU-Net was included as a baseline, only its 2D implementation was evaluated under our unified experimental protocol. This baseline therefore represents a custom adaptation that may underestimate the true capability of the nnU-Net framework. The complete self-configuring nnU-Net pipeline and whole-volume 3D evaluation were not explored, which should be addressed in future work. Likewise, modern ViT baselines were not included in our comparative experiments, and a direct experimental comparison against Transformer-based architectures remains future work. In addition, because of the retrospective design and limited dataset scale, the reported results should be interpreted as indicating promising accuracy for research and assistive purposes rather than as evidence of validated clinical utility. The private test set included only 22 patients, which may limit the statistical power to detect significant performance differences between methods. The heterogeneous nature of renal tumors across stages and histological subtypes, combined with the single-institution data source, may limit the generalizability of the current findings. Prospective multicenter studies with larger and more diverse patient cohorts are needed before conclusions regarding clinical deployment can be made with confidence.

In summary, this study proposed GAM-DeepLabV3+, a CT renal tumor segmentation model built on an improved DeepLabV3+ framework incorporating a lightweight MobileNetV2 encoder, an ASPP module, and a GAM attention mechanism in the decoder. The model demonstrated promising segmentation accuracy on both the private dataset (Dice: mean 0.939, SD 0.008; IoU: mean 0.928, SD 0.012; HD95: mean 1.485, SD 0.522 pixels) and the public KiTS19 dataset (Dice: mean 0.928, SD 0.006; IoU: mean 0.902, SD 0.008; HD95: mean 2.536, SD 0.742 pixels), while offering improved inference speed relative to standard DeepLabV3+. By integrating lightweight backbone design with global-local attention feature fusion, the proposed approach is designed to address several challenges in renal tumor CT segmentation, such as blurred tumor boundaries and class imbalance; a dedicated, tumor size–stratified evaluation of small-lesion performance remains an avenue for future work. These findings support the potential of GAM-DeepLabV3+ as a research tool and a basis for further development toward clinically assistive automated segmentation systems. Prospective validation on larger, multicenter datasets will be an essential next step in assessing broader clinical applicability.