Automated Renal Tumor Segmentation in Computed Tomography Images Using a Global Attention-Based DeepLabV3+ Model: Algorithm Development and Validation.

BACKGROUND

The rising global incidence of renal tumors necessitates precise diagnostic interventions. Accurate segmentation of computed tomography (CT) scans is essential for nephron-sparing surgery and radiotherapy. However, conventional manual delineation is labor-intensive and prone to significant interobserver variability due to tumor morphological heterogeneity. There is an urgent clinical demand for robust, automated segmentation solutions.

OBJECTIVE

This study aims to develop and validate GAM-DeepLabV3+, an automated framework designed to address boundary ambiguity and high false-positive rates in complex renal imaging scenarios.

METHODS

We propose an optimized encoder-decoder architecture specifically tailored for renal mass detection. The framework incorporates three key innovations: (1) a lightweight MobileNetV2 backbone to minimize computational overhead for clinical deployment; (2) an Atrous Spatial Pyramid Pooling (ASPP) module to capture multiscale contextual information; and (3) a Global Attention Mechanism (GAM) in the decoder to enhance channel-spatial interactions, thereby refining boundary delineation by suppressing background noise. The model was rigorously evaluated on a private clinical dataset (n=218) and the KiTS19 benchmark (n=210).

RESULTS

GAM-DeepLabV3+ consistently outperformed state-of-the-art baselines. On the private dataset, the model achieved a mean Dice similarity coefficient (DSC) of 0.939 (SD 0.008), significantly surpassing feature pyramid network (FPN; mean 0.893, SD 0.008; P<.001) and no-new-Net (nnU-Net; 2D, custom; mean 0.908, SD 0.013; P<.001). It also achieved a mean 95% Hausdorff distance (HD95) of 1.485 (SD 0.522) pixels. On the KiTS19 dataset, it maintained a mean robust DSC of 0.928 (SD 0.006). To facilitate clinical translation, a demonstration-only online platform was developed.

CONCLUSIONS

The GAM-DeepLabV3+ framework provides an accurate, efficient, and fully automated solution for renal tumor segmentation. By overcoming boundary ambiguity and optimizing feature fusion, this approach shows potential as a decision-support aid, pending future validation with 3D reconstruction.