JMIR Publications

ABSTRACT

Background:

Renal tumors represent one of the most common malignancies worldwide, with incidence rates continuing to rise. Early detection and precise treatment are crucial for effective disease management. Accurate segmentation of renal tumors in CT images plays a critical role in lesion localization and radiotherapy planning. However, current segmentation methods largely depend on manual delineation by radiologists, which is both time-consuming and subject to inter-observer variability due to tumor heterogeneity, posing significant challenges for automation.

Objective:

This study aims to develop and validate an automated renal tumor segmentation algorithm that addresses the challenges of blurred tumor boundaries and false positives in CT imaging, thereby enhancing clinical decision-making and treatment planning.

Methods:

We propose GAM-DeepLabV3+, an automatic segmentation model built upon the DeepLabV3+ encoder-decoder framework. The architecture integrates three key innovations. First, an enhanced MobileNetV2 backbone combined with a spatial pyramid pooling layer is employed to extract comprehensive low-level features and critical tumor information from CT scans. Second, a Global Attention Mechanism (GAM) module is incorporated into the decoder to efficiently fuse deep and shallow features, improving boundary delineation. Third, multi-scale feature integration enables the network to adaptively focus on tumor regions with varying sizes and shapes. The model was trained and evaluated on both a private dataset and the publicly available KiTS19 dataset.

Results:

Experimental evaluation demonstrates that GAM-DeepLabV3+ achieves superior segmentation performance, with Dice coefficients of 0.92 on the private dataset and 0.98 on the KiTS19 dataset. These results significantly outperform conventional methods, particularly in cases with complex tumor morphology. Furthermore, we developed a freely accessible online platform (http://www.cppdd.cn/KAI) to facilitate clinical application and support preoperative planning.

Conclusions:

The proposed GAM-DeepLabV3+ model provides accurate, efficient, and fully automated renal tumor segmentation, reducing dependence on manual annotation while maintaining clinical-grade precision. By addressing key challenges in renal tumor imaging, our approach offers valuable support for surgical planning and treatment, and holds promise for broader integration into clinical workflows.