GennUNet - Abdominal Organ Segmentation

Model Description

GennUNet is a medical image segmentation model for computed tomography (CT) scans. Built on the nnUNet architecture, it achieves high generalizability across diverse datasets by leveraging a unified dataset from BTCV, AMOS, and TotalSegmentator. The model is optimized to handle variations in imaging properties, demographics, and anatomical features, making it robust for real-world clinical applications.

Model Details

• Developed by: Nicolás Álvarez Llopis
• Supervised by: María de la Iglesia Vayá, Dario García Gasulla
• Institution: Universitat Politècnica de Catalunya (UPC), Universitat de Barcelona (UB), Universitat Rovira i Virgili (URV)
• License: Apache 2.0
• Architecture: nnUNet (Fully Convolutional Network)
• Domain: Medical Image Segmentation
• Modality: Computed Tomography (CT)
• Tasks: Abdominal Organ Segmentation
• Training Framework: PyTorch, MONAI
• Link to Repositories: Code, Datasets

Datasets

GennUNet was trained using a unified dataset consisting of three large-scale abdominal organ segmentation datasets:

• BTCV (Beyond the Cranial Vault)
• AMOS (Abdominal Multi-Organ Segmentation)
• TotalSegmentator

The datasets were processed to remove redundant and inconsistent samples, including intensity normalization, orientation normalization, foreground cropping, and spacing standardization to ensure consistent training input.

Evaluation Results

GennUNet was evaluated using a five-fold cross-validation approach, demonstrating superior segmentation performance:

Organ	Dice Score (%)
Spleen	97.4
Right Kidney	96.5
Left Kidney	96.4
Gallbladder	86.8
Esophagus	89.0
Liver	98.2
Stomach	94.2
Aorta	96.6
Inferior vena cava	93.1
Pancreas	89.4
Right adrenal gland	84.9
Left adrenal gland	85.2

GennUNet demonstrates strong generalization across datasets, outperforming transformer-based models and previous state-of-the-art segmentation models.

Training Details

• Loss Function: Dice Loss + Cross-Entropy Loss
• Optimizer: Adam + Polymonial Learning Rate scheduler
• Initial Learning Rate: 0.01
• Batch Size: 2
• Augmentation: Rotation, scaling, Gaussian noise, contrast adjustment, mirroring
• Training Duration: 1000 epochs

Intended Use

This model is designed for:

• Automated segmentation of abdominal organs in CT scans
• Assisting radiologists in diagnostic workflows
• Medical research involving organ volumetry and disease characterization

Limitations & Challenges

• Small Organ Segmentation: Lower Dice scores for smaller organs like adrenal glands (85.2%) due to limited visual information.
• Contrast Variability: Variations in CT contrast levels affect model performance, particularly in datasets with mixed contrast phases.
• Data Bias: Reliance on public datasets introduces biases that may not represent global diversity.

Future Work

• Fine-tuning on additional datasets to enhance generalizability
• Exploration of transformer-based enhancements
• Incorporation of domain adaptation techniques to mitigate dataset bias

Citation

If you use GennUNet in your research, please cite:

@mastersthesis{alvarez2024diverse,
  title={From diverse CT scans to generalization: towards robust abdominal organ segmentation},
  author={{\'A}lvarez Llopis, Nicol{\'a}s},
  year={2024},
  school={Universitat Polit{\`e}cnica de Catalunya}
}

---