An Efficient Multi-Scale Hybrid Transformer for High-Precision Renal Malignancy Identification

Abstract

Renal cell carcinoma remains a critical challenge in global oncology, where the prognosis is heavily dependent on early-stage intervention. While Computed Tomography (CT) is the gold standard for diagnosis, the automated localization of tumors is often hindered by complex anatomical backgrounds and significant variations in lesion scale. To address these limitations, this research proposes An efficient  Multi-Scale Hybrid  Transformer Network for high precision renal malignancy identification. To mitigate the constraints of limited clinical datasets, the framework utilizes an intensive data augmentation pipeline. The architecture integrates a specialized multi-scale transformer backbone designed to capture long-range spatial dependencies and handle morphological diversity in tumors. A key innovation is the inclusion of a lightweight channel-wise attention module following the feature fusion stage, which selectively excites relevant feature maps and suppresses noise, thereby refining the precision of disease delineation. Experimental results indicate that the proposed model provides a computationally efficient solution for clinical environments, achieving Intersection over Union (IoU) of 97.28%, a Precision of 96.11%, a Recall of 95.12%, an F1-Score of 96.11%, and a mean Average Precision (mAP) of 96.56%. These metrics demonstrate that the proposed hybrid approach outperforms traditional convolutional models, offering a robust tool for real-time diagnostic support in radiology