STC-UNet: renal tumor segmentation based on enhanced feature extraction at different network levels

Area of Science:

• Medical Image Analysis
• Deep Learning in Urology
• Computational Pathology

Background:

• Accurate renal tumor segmentation is vital for diagnosis and treatment but challenged by indistinct boundaries and morphological variations.
• Existing deep learning models struggle with extracting specific renal tumor features across different network hierarchies, limiting segmentation accuracy.
• There is a need for specialized models that can effectively capture multi-scale and contextual features of renal tumors.

Purpose of the Study:

• To propose a novel deep learning architecture, the Selective Kernel, Vision Transformer, and Coordinate Attention Enhanced U-Net (STC-UNet), for improved renal tumor segmentation.
• To enhance the extraction of multi-scale and long-range contextual features specific to renal tumors.
• To improve the localization and recovery of tumor regions through enhanced feature representation.

Main Methods:

• Integration of Selective Kernel modules in shallow layers to capture multi-scale detailed features.
• Incorporation of non-patch Vision Transformer modules in deeper layers for global contextual information and fine-grained feature extraction.
• Implementation of Coordinate Attention modules in the decoder for enhanced feature recovery and precise tumor localization.

Main Results:

• The STC-UNet model demonstrated significant improvements over the baseline model on the KiTS19 dataset.
• Performance gains included increases in IoU (1.60%), Dice (2.02%), Accuracy (2.27%), Precision (1.18%), Recall (1.52%), and F1-score (1.35%).
• The proposed method outperformed other advanced algorithms in both visual quality and objective evaluation metrics.

Conclusions:

• The STC-UNet effectively addresses the limitations of existing models in renal tumor segmentation by enhancing feature extraction capabilities.
• The combined use of Selective Kernel, Vision Transformer, and Coordinate Attention modules leads to superior segmentation performance.
• STC-UNet shows promise for clinical application, offering more accurate segmentation to aid in renal tumor diagnosis and treatment planning.