DeepOrgan: Multi-level Deep Convolutional Networks for Automated Pancreas Segmentation

Holger R. Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B. Turkbey, Ronald M. Summers

Introduction

Automated pancreas segmentation in Computed Tomography (CT) scans presents a significant challenge in medical imaging. Unlike organs with consistent morphology like the liver, the pancreas exhibits high anatomical variability in size, shape, and orientation. Furthermore, its boundaries are often obscured by surrounding visceral fat of similar intensity, leading to "fuzzy" edges that traditional "top-down" atlas-based methods struggle to delineate accurately. DeepOrgan addresses these limitations by introducing a robust, "bottom-up" hierarchical approach using Deep Convolutional Networks (ConvNets).

Methodology

The proposed system utilizes a multi-level pipeline that transitions from local pixel analysis to regional context to achieve precise organ segmentation.

Hierarchical pipeline for pancreas segmentation, from candidate region generation to final probability map.

1. Candidate Region Generation

The process begins by over-segmenting the CT image into superpixels. A fast Random Forest classifier then filters these regions, retaining any that potentially contain pancreatic tissue. This stage prioritizes high recall (>97%) to ensure the organ is not prematurely excluded.

2. P-ConvNet (Local Patches)

A dedicated ConvNet (P-ConvNet) analyzes "2.5D" patches—orthogonal axial, coronal, and sagittal slices centered on a voxel. This step generates a dense probability map indicating the likelihood of each voxel belonging to the pancreas.

Architecture of the Deep Convolutional Network (ConvNet) used for patch and regional analysis.

3. R-ConvNets (Regional Context)

The core innovation lies in the Regional ConvNets (R-ConvNets), which evaluate superpixel candidates across multiple scales. This includes an R1-ConvNet for raw CT intensities and a stacked R2-ConvNet that processes both intensities and the probability maps from the P-ConvNet. A "zoom-out" strategy samples bounding boxes at various scales, providing the network with both high-resolution local detail and broader anatomical context necessary for disambiguating the pancreas from surrounding structures.

4. Post-processing

To ensure spatial consistency, the final output is refined using 3D Gaussian smoothing and 2D Conditional Random Fields (CRF) to sharpen the segmented boundaries.

Data and Evaluation

The model was developed and validated using a dataset of 82 contrast-enhanced abdominal CT volumes, with ground truth provided by expert manual tracings. Performance was evaluated via 4-fold cross-validation to ensure generalizability.

Key Insights

  • Contextual Awareness: The "zoom-out" feature allows the network to recognize surrounding organs and fat, providing the necessary context to identify the variable pancreas.
  • Bottom-Up Robustness: By constructing the segmentation from local and regional components rather than fitting a rigid global template, the system handles extreme anatomical variability more effectively.
  • Data Augmentation: To mitigate the small dataset size, Thin-Plate-Spline (TPS) deformations were used to "warp" training images, effectively increasing data diversity and preventing overfitting.
  • Computational Efficiency: While training is intensive (approx. 55 hours), inference for a new patient takes only 1 to 3 minutes, a significant improvement over traditional methods.

Results

DeepOrgan achieved a state-of-the-art Dice Similarity Coefficient (DSC) of 71.8% ± 10.7% at the time of its 2015 publication. The system demonstrated high stability and a significant performance leap from the initial candidate stage (26.1% DSC). Qualitatively, the model successfully segmented 80 out of 82 cases with over 50% accuracy, with failures primarily occurring due to over-segmentation into adjacent organs.

Limitations and Future Work

Current limitations include the reliance on 2D/2.5D processing due to the computational cost of 3D "supervoxels" at the time. Additionally, a performance gap between training (83.6%) and testing (71.8%) suggests a need for larger, more diverse datasets. Future research should also explore the model's performance on pathological cases, such as those involving pancreatic tumors, which further complicate the organ's appearance.

Summary updated

BibTeX
@inproceedings{roth2015deeporgan,
  author = {Holger R. Roth and Le Lu and Amal Farag and Hoo-Chang Shin and Jiamin Liu and Evrim B. Turkbey and Ronald M. Summers},
  title = {{DeepOrgan: Multi-level Deep Convolutional Networks for Automated Pancreas Segmentation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2015},
  doi = {10.1007/978-3-319-24553-9\_68},
}
Figure