MDM-U-Net: A novel network for renal cancer structure segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of renal cancer structures in contrast-enhanced computed tomography (CT) images is crucial for diagnosis and treatment planning.
• Existing segmentation methods may face challenges in capturing complex anatomical details and variations in renal cancer structures.

Purpose of the Study:

• To propose and evaluate a novel encoder-decoder segmentation network, MDM-U-Net, for precise segmentation of renal cancer structures (kidney, tumors, vessels).
• To enhance feature extraction, information exploitation, and segmentation performance through specialized network components.

Main Methods:

• Developed MDM-U-Net, incorporating a multi-scale anisotropic convolution block for feature extraction, a dual activation attention block for information guidance, and a multi-scale deep supervision mechanism for decoder layer supervision.
• Trained and validated the model on the public KiPA22 dataset and an in-house dataset.

Main Results:

• MDM-U-Net achieved state-of-the-art (SOTA) performance on the KiPA22 dataset, ranking first in 6 out of 12 evaluation metrics.
• On the in-house dataset, the model demonstrated superior generalization, achieving SOTA performance in 9 out of 12 metrics.

Conclusions:

• The proposed MDM-U-Net is a feasible and generalizable model for renal cancer structure segmentation from contrast-enhanced CT scans.
• The network's architecture effectively improves segmentation accuracy, outperforming existing methods and showing significant clinical potential.