Link to paper
The full paper is available here.
You can also find the paper on PapersWithCode here.
Abstract
- Demonstrated automated segmentation of ccRCC, cysts, and normal kidney parenchyma in VHL patients using CNN on MRI
- Queried 115 VHL patients and 117 scans with 504 ccRCCs and 1171 cysts from 2015 to 2021
- Evaluated U-Net performance on 10 randomized splits of the cohort using DSC
- 2D U-Net achieved an average ccRCC lesion detection AUC of 0.88 and DSC scores of 0.78, 0.40, and 0.46 for kidney, cysts, and tumors, respectively
- 3D U-Net achieved an average ccRCC lesion detection AUC of 0.79 and DSC scores of 0.67, 0.32, and 0.34 for kidney, cysts, and tumors, respectively
Paper Content
Introduction
- VHL syndrome is an inherited disease that can cause clear cell renal cell carcinoma (ccRCC)
- ccRCC is the leading cause of mortality in VHL patients
- MRI is the imaging method of choice for lifelong surveillance
- Radiomics can be used to predict tumor behavior and individualize treatment
- Automated segmentation of renal lesions is needed to improve radiomics workflow
- Recent research on automated RCC detection and segmentation uses CT scans
- Deep convolutional neural networks (CNN) are used to segment kidney and RCC tumors
- No published study has built a deep learning model based on MRI scans of VHL patients
Material and methods
Patient cohort
- Registry of VHL patients who underwent renal surgery between 2015-2021
- Written consent obtained
- Ethical guidelines followed
- MRIs captured before surgery
- MRI examinations conducted on four MRI scanners
- Routine imaging sequences used
- Gadobutrol administered
- Kidney with pathology segmented
- Smaller cysts < 10mm excluded from segmentation
U-net
- U-Net was trained to segment kidney, cysts, and RCC tumors in 2D axial slices
- Leaky ReLU and batch normalizations were used
- Unknown pathology was excluded using an “ignore label”
- Intensity thresholding was used to mask out unannotated cysts
- U-Net was trained to minimize binary crossentropy using Adam stochastic gradient descent
- Horizontal image flipping was used for data augmentation
- 3D U-Net used 32x32x32 patches with a stride of 16x16x16 for training and 28x28x28 for inference
Data preprocessing
- MR scans were identified, registered and resampled to have 1mm x 1mm axial slices.
- 3D U-Net models were resampled to 1mm x 1mm x 3mm.
- Images were registered using niftyreg.
- MRIs were normalized using Rician normalization.
Statistical analysis
- Evaluated U-Net performance on 10 randomized splits of patients
- 75% of data used for training, 25% for testing
- 10% of training data used for validation
- Dice similarity coefficient (DSC) used to pick best model state
- Evenly sampled from list of easy and complex cases
- Measured average per-tumor DSC
- Also measured AUC of ccRCC tumor detection
- Calculated standard deviation for AUC
Computing environment
- Experiments and analysis were run on Linux (CentOS 7)
- Linux (CentOS 7) was used as the operating system
Results
- Used image database of 115 individuals and 117 scans with 504 pathologically confirmed ccRCCs and 1171 cystic lesions
- Average number of ccRCC lesions per patient was 4.3 ± 3.1
- U-Net was able to detect tumors with relatively high AUC (0.89)
- Model accurately detected nearly 100% of the background in all splits
- 3D-U-Net performance was lower than 2D U-Net, but no significant difference between the two models’ performances
Kidney
Discussion
- Aim of study: Determine if CNN can be used to identify and segment ccRCC tumors from cysts and renal parenchyma in VHL patients
- Results: High accuracy in detecting renal tumors on MRI (AUC 0.79-0.88) and showed promise in segmentation (DSC 0.34-0.46)
- Artificial intelligence may help orient radiologists to lesions, allowing for prompt and appropriate management
Comparing the performance
- Examined prior studies on CT scans
- Employed continual learning curve to bring automatic segmentation of renal cell carcinoma to current stage
- Majority of models in KITs-19 challenge demonstrated encouraging performance
- Groundwork for future MRI autosegmentation models
- MRI imaging has several challenges compared to CT
- No other peer-reviewed research has attempted to segment ccRCC lesions and cysts using either 2D or 3D U-Net
- Model demonstrated improvement in tumor segmentation compared to mentioned study
- Results for kidney parenchyma from model may differ from other studies due to differences in ground truth definition
- Model achieved defensible DSC
- Range in cystic forms and morphologies can complicate machine detection and segmentation of cysts
- Presence of many unannotated structures can be challenging
Limitation
- Study has small sample size
- Underrepresented tumors degraded model’s performance
- Model built on four contrasts enhanced T1 sequences
- Improved performance expected with addition of other MR sequences
- Coregistration of images challenging part of adding sequences
Conclusion
- Demonstrated encouraging findings for 2D and 3D U-Net detection and segmentation of ccRCC tumors on MRI
- Wide variety of pathologies and appearances necessitates data set development
- U-Net generalizes better when training set is representative
- 2D U-Net can access more global image context, 3D U-Net is more localized
- Automated hyperparameter tuning of both 2D and 3D U-Net proposed for optimal performance