Accurate Kidney Segmentation in CT Scans Using Deep Transfer Learning.

A competitive model for kidney segmentation in CT scans is trained using the publicly-available KiTS19 dataset. The model performed well against the KiTS19 test dataset, achieving a Sorensen–Dice coefficient of 0.9620 when generating kidney segmentation masks from CT scans. The algorithm employed is U-Net, a common tool used to segment biomedical images of various modalities, including MRI and CT scans. The model is trained using nnU-Net, an open-source framework for training U-Net. To help bring Deep Learning to kidney stone diagnosis and treatment, this promising model is then applied to a dataset developed by the research team, comprised of CT scans from patients who underwent treatment for kidney stones between 2011 and 2014. Despite overall success, the model appears sensitive to changes in features between the two datasets, with some segmentation masks working very well and others unable to correctly separate the kidney from the surrounding anatomy. Improving this model further will enable advanced research in deep learning tools to aid urologists’ decision making for best procedure outcomes.