Global Attention DeepLabV3+ Achieves High Dice Score for Renal Tumor CT Segmentation

Automated segmentation model GAM-DeepLabV3+ achieved a mean Dice similarity coefficient of 0.939 for renal tumors on CT, outperforming FPN and nnU-Net (P<.001), with a 95% Hausdorff distance of 1.485 pixels. External validation on KiTS19 yielded Dice 0.928.

• GAM-DeepLabV3+ achieved a mean Dice of 0.939 (SD 0.008) on the private dataset, significantly outperforming FPN (0.893) and nnU-Net (0.908), both P<.001.
• On the external KiTS19 benchmark, it maintained a robust Dice of 0.928 (SD 0.006).
• The study is limited to 2D segmentation; validation with 3D reconstruction is pending.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.