Cascaded Regression Neural Nets for Kidney Localization and Segmentation-free Volume Estimation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nephrology

Background:

• Kidney volume is a crucial biomarker for diagnosing kidney diseases like chronic kidney disease.
• Current kidney volume estimation methods often require a time-consuming kidney segmentation step.
• Accurate kidney localization in volumetric medical images is essential for subsequent analysis.

Purpose of the Study:

• To develop an integrated deep learning approach for kidney localization and segmentation-free renal volume estimation.
• To improve the efficiency and accuracy of kidney volume assessment in computed tomography (CT) scans.

Main Methods:

• A selection-convolutional neural network was employed for kidney localization along the axial direction.
• A combined sagittal-axial Mask R-CNN was utilized to generate a 3D organ bounding box.
• A fully convolutional network was used for direct kidney volume estimation, bypassing segmentation.

Main Results:

• The method achieved a kidney boundary wall localization error of approximately 2.4mm.
• The mean kidney volume estimation error was approximately 5%.
• Validation was performed on CT scans from 100 patients and the 2019 Kidney Tumor Segmentation Challenge database.

Conclusions:

• The proposed integrated deep learning approach enables accurate kidney localization and segmentation-free volume estimation.
• This method offers a more efficient alternative to traditional techniques, potentially enhancing clinical workflows for kidney disease management.