Deep learning kidney segmentation with very limited training data using a cascaded convolution neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nephrology

Background:

• Deep learning for medical image segmentation necessitates extensive annotated datasets.
• Manual image annotation is labor-intensive, creating a bottleneck for clinical applications.
• This limits the availability of ground truth data, especially for specialized imaging like kidney segmentation.

Purpose of the Study:

• To assess the feasibility of kidney segmentation using deep learning convolutional neural network (CNN) models.
• To investigate the effectiveness of training CNN models with a minimal number of subjects (few-shot learning).
• To evaluate a 3D augmentation strategy for enhancing segmentation performance with limited data.

Main Methods:

• Employed a few-shot deep learning strategy with 3D augmentation on T1-weighted MR images.
• Trained cascaded CNN models using data from one, three, and six subjects for segmentation.
• Evaluated model performance using Dice and Jaccard coefficients on independent test cohorts.

Main Results:

• Achieved a mean Dice coefficient of 0.85 with only one training subject and 0.91 with six.
• The cascaded network significantly improved segmentation over a single U-Net, particularly with limited training data (e.g., Dice 0.835 vs. 0.759 with one subject).
• Demonstrated superior performance in a second independent cohort, highlighting robustness.

Conclusions:

• Few-shot kidney segmentation using 3D augmentation is feasible and effective.
• Cascaded CNN architectures enhance segmentation accuracy compared to single models when training data is scarce.
• This approach offers a viable solution for medical image segmentation challenges with limited ground truth masks.