Segmentation of kidney mass using AgDenseU-Net 2.5D model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• The Kidney and Kidney Tumor Segmentation Challenge 2021 (KiTS21) dataset aids research in kidney mass segmentation.
• Accurate segmentation of kidneys, tumors, and cysts is crucial for assessing kidney mass complexity and aggressiveness.
• Traditional 3D deep learning models require significant computational resources.

Purpose of the Study:

• To propose a computationally efficient method for segmenting kidneys, tumors, and cysts.
• To evaluate the performance of the proposed method on the KiTS21 dataset.
• To provide a reference for kidney tumor segmentation techniques.

Main Methods:

• A two-step approach involving automatic down-sampling of 3D CT images to reduce volume while preserving features.
• Fine segmentation using the AgDenseU-Net (Attention gate DenseU-Net) 2.5D model.
• Evaluation using Hierarchical Evaluation Classes (HECs) specific to KiTS21, alongside standard metrics.

Main Results:

• The AgDenseU-Net model achieved high Dice scores on the KiTS21 dataset: 0.971 for kidney, 0.883 for kidney mass, and 0.815 for tumor.
• Without HECs, the model obtained Dice scores of 0.950 for kidney, 0.878 for tumor, and 0.746 for cyst.
• The proposed method demonstrates effectiveness in segmenting kidney, tumor, and cyst components.

Conclusions:

• The proposed two-step deep learning scheme significantly reduces computational resource requirements for kidney mass segmentation.
• The AgDenseU-Net model shows promising results for accurate segmentation of kidneys, tumors, and cysts.
• This method serves as a valuable reference for future kidney tumor segmentation research and clinical applications.