https://atlas.rsna.org/schemas/2025-11/model.json

Style Transfer–assisted Deep Learning Method for Kidney Segmentation at Multiphase MRI

https://dx.doi.org/10.1148/ryai.230043

1.0

Corresponding author: Ivan Pedrosa, MD, PhD, University of Texas Southwestern Medical Center, 2201 Inwood Rd, Suite 202, Dallas, TX 75390-9085.

Partially funded by the National Institutes of Health (grant R01CA154475). Computational resources provided by the BioHPC supercomputing facility, Lyda Hill Department of Bioinformatics, UT Southwestern Medical Center.

Retrospective, HIPAA-compliant, institutional review board–approved study with waiver of informed consent.

Semisupervised pipeline combining CycleGAN (PyTorch) for style transfer to synthetic MCE MRI and Mask R-CNN (TensorFlow, InceptionResNetV2 backbone; segmentation head only) for kidney segmentation.

CycleGAN implementation: https://github.com/junyanz/CycleGAN; TensorFlow Object Detection (Mask R-CNN): https://github.com/tensorflow/models/tree/master/research/object_detection. Trained study models not publicly released.

Automates kidney segmentation on multiphase contrast-enhanced MRI, reducing manual annotation burden and enabling downstream quantitative analyses and radiomics workflows.

Workflow optimization (automated image segmentation to support analysis of mpMRI examinations).

Fully automatic segmentation at inference; semisupervised training using synthetic images generated from T2-weighted images.

Automated segmentation of kidneys in coronal multiphase contrast-enhanced (precontrast, corticomedullary, early nephrographic, nephrographic) MRI examinations in patients with renal masses, in a research setting.

Coronal MCE T1-weighted MRI acquisitions across four phases (precontrast, corticomedullary, early nephrographic, nephrographic). T2-weighted coronal images and kidney masks used for style transfer training and as ground truth during training on synthetic images.

1) Train CycleGAN to generate corticomedullary-style images from T2W (Cohort 1). 2) Train additional CycleGANs to generate precontrast, early nephrographic, and nephrographic images from corticomedullary images (Cohort 2). 3) Train four phase-specific Mask R-CNN models on synthetic phase images using T2W kidney masks as ground truth. 4) Apply inference to original MCE MRI phases; perform 3D morphologic postprocessing (dilation, erosion, clean, majority, fill).

Small cohorts; validation limited to coronal-plane MCE MRI; generalizability to axial acquisitions and other institutions not established; focuses on kidney segmentation only (renal masses not segmented); training and evaluation limited to patients with renal masses.

Research use for automated kidney segmentation in MCE MRI; authors suggest future extension to renal mass segmentation and other mpMRI acquisitions.

Implementation details: CycleGAN (PyTorch); Mask R-CNN (TensorFlow, InceptionResNetV2); grayscale input, 512×512; batch size 4; 100 epochs; initial learning rate 0.008; momentum 0.9; gradient clipping norm 10; pretrained weights (COCO); real-time flips for augmentation; model selection by minimum validation loss; hardware: Nvidia Titan V100 32 GB GPU. N4 bias field correction applied. 3D morphologic postprocessing applied after 2D inference.