Related Experiment Video
Updated: Aug 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Deep learning-based decision forest for hereditary clear cell renal cell carcinoma segmentation on MRI
Nathan Lay1, Pouria Yazdian Anari2, Aditi Chaurasia2
1Artificial Intelligence Resource, Molecular Imaging Branch, National Cancer Institute, Bethesda, Maryland, USA.
Medical Physics
|March 1, 2023
Summary
A novel hinge forest (HF) neural network improved segmentation of clear cell renal cell carcinoma (ccRCC) tumors in MRI scans for von Hippel-Lindau syndrome (VHL) patients. This AI approach offers better tumor characterization and prediction than U-Net.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- von Hippel-Lindau syndrome (VHL) patients have a high risk of developing clear cell renal cell carcinoma (ccRCC).
- Lifelong MRI surveillance is crucial for VHL patients to monitor disease progression and avoid radiation exposure.
- Accurate segmentation of kidney and tumor structures on MRI aids in lesion characterization, volumetric analysis, and growth prediction.
Purpose of the Study:
- To develop and evaluate a novel AI-based segmentation methodology for ccRCC on various MRI contrast phases.
- To compare the performance of a new hinge forest (HF) neural network against the U-Net architecture for ccRCC segmentation.
Main Methods:
- A novel differentiable hinge forest (HF) neural network was applied for segmenting kidney parenchyma, cysts, and ccRCC tumors.
- The HF model was trained and tested on a large dataset of 117 MRI scans from 115 VHL patients, encompassing 504 ccRCCs and 1171 cysts.
- Performance was evaluated against U-Net using the Dice similarity coefficient (DSC) on randomized data splits.
Main Results:
- The HF model achieved competitive segmentation performance, with specific strengths in tumor segmentation.
- While U-Net showed slightly better performance for kidney parenchyma segmentation (DSC 0.78 vs. 0.75), HF demonstrated superior tumor segmentation (DSC 0.53 vs. 0.46).
- Statistical analysis indicated significant differences in performance for both kidney parenchyma and tumor segmentation between the two methods.
Conclusions:
- The hinge forest (HF) approach shows promise for improved ccRCC segmentation in VHL patients compared to U-Net.
- The interpretability of the HF method, through its decision tree structure, may offer insights for automated tumor characterization.
- This study highlights the potential of advanced deep learning techniques for enhancing diagnostic accuracy in VHL syndrome management.
