Automatic renal mass segmentation and classification on CT images based on 3D U-Net and ResNet algorithms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate evaluation of renal masses in CT images is crucial for patient management.
• Manual segmentation and classification of renal lesions can be time-consuming and subjective.
• Deep learning offers potential for automated analysis of medical imaging data.

Purpose of the Study:

• To develop and validate a cascade deep learning method for automated renal mass segmentation and classification in CT images.
• To utilize 3D U-Net for kidney boundary segmentation and renal mass detection/segmentation.
• To employ ResNet for the classification of segmented renal masses.

Main Methods:

• A cascade deep learning architecture combining 3D U-Net and ResNet was developed.
• 3D U-Net was used for initial kidney boundary segmentation to define regions of interest.
• An ensemble 3D U-Net model performed mass detection and segmentation, followed by ResNet classification.

Main Results:

• The algorithm achieved a Dice Similarity Coefficient (DSC) of 0.99 for kidney boundary segmentation.
• Average DSC for renal mass delineation was 0.75 and 0.83.
• Recall for renal mass detection was 84.54% and 75.90%, with classification accuracy of 86.05% (<5 mm) and 91.97% (≥5 mm).

Conclusions:

• A fully automated deep learning-based method for renal mass segmentation and classification in CT images was successfully developed.
• The algorithm demonstrated accurate localization and classification capabilities for renal masses.
• This automated approach holds promise for improving the efficiency and accuracy of renal mass evaluation.